DEV Community: Manohar Papasani

Qodo tops independent AI code review tests, but GitHub Copilot wins on integration.

Manohar Papasani — Mon, 09 Nov 2026 19:30:59 +0000

The top pick is not the easiest tool to plug into your workflow.

LogRocket ran the same tests on five AI code review tools and ranked the results.
Accuracy and speed set the tiers; integration is judged separately.

Does it find the problems that matter?
Is it fast and detailed enough for daily use?
Does it fit the tools your team already uses?

🔴 S TIER
Qodo

Fastest and most detailed tool in LogRocket's test of five AI code review tools.
Its prompt-based workflow makes it easy to dig into exactly what you need.

🟠 A TIER
Traycer

Took second place in LogRocket's test, but was not as quick or thorough as Qodo.
Still gives useful insights and is simple to work with.

⏳ NOT RANKED YET
GitHub Copilot

LogRocket rates it the easiest to integrate, since most teams already work with GitHub.
The source gives it no accuracy tier, so it is not placed above or below the others here.

What this means for your team:

Trial Qodo first if speed and detail matter most.
If your code lives on GitHub, weigh Copilot's easy setup against Qodo's stronger results.
Run a shortlist against your own real bugs before committing.

Pick one tool from the S tier and run it on your next pull request this week.

P.S. Vendor-published benchmarks, such as Kodus's 38-bug test, are left out of the ranking because the vendor published them.

Originally shared on LinkedIn.

An AI interview question every day, leadership for tech people and hands-on AI learning:

Seven things happened in AI today; the one worth your attention is Clinics On Cloud.

Manohar Papasani — Mon, 09 Nov 2026 12:01:03 +0000

Here is the rundown.

💰 1. Clinics On Cloud: Clinics On Cloud reportedly raised $12.5 million in a Series A round from Dubai‑based investors via Model42, to bolster its AI‑driven healthcare tools, expand R&D and scale its microclinic network (reported).
https://entrepreneur.economictimes.indiatimes.com/news/funding/funding-wrap-clinics-on-cloud-peppermint-robotics-giva-raise-fresh-capital/135600777

💰 2. Lazarus AI: Lazarus AI reportedly raised $43.0M in a round, bringing its total equity funding to $49M (reported).
https://www.alleywatch.com/2026/11/the-weekly-notable-startup-funding-report-11-9-26

🚀 3. GPT-6 Cyber: OpenAI said GPT‑6 Cyber will initially be offered via its API and Codex to a limited set of approved Daybreak Red customers (reported).
https://www.theregister.com/security/2026/11/05/openai-releases-gpt-6-cyber-and-promises-this-one-stays-within-rules-of-engagement/5307855

💰 4. Kalepa: Kalepa reportedly raised $75M in a Series B round led by Updata Partners, with IA Ventures and Inspired Capital Partners participating, bringing its total reported equity funding to $91M (reported).
https://alleywatch.com/2026/11/the-new-york-startup-daily-funding-report-11-5-2026

💰 5. Trajectory: Trajectory said it closed a $45 million Series A, valuing the 18‑person firm at about $300 million, with Sequoia leading and Nvidia, Bessemer Venture Partners and Bain Capital Ventures participating (announced).
https://techcrunch.com/2026/11/02/trajectory-raises-45m-to-make-ai-model-training-accessible-to-nontechnical-experts

💰 6. Traversal: Traversal reportedly raised $120 million, valuing the startup at about $1 billion (reported).
https://americanbazaaronline.com/2026/11/08/traversal-120-million-funding-billion-dollar-valuation-491804

🚀 7. Gemini 4 Flash: Google has not announced Gemini 4 Flash as of November 8, 2026 (confirmed), and two X accounts estimate a release around November 18, 2026 (reported).
https://cellcog.ai/blog/gemini-4-flash-release-date

💾 Save this if you are tracking AI releases and funding.
Daily AI digest, with a source link under every item.

Originally shared on LinkedIn.

An AI interview question every day, leadership for tech people and hands-on AI learning:

How do you evaluate the quality and accuracy of responses generated by an LLM?

Manohar Papasani — Mon, 09 Nov 2026 08:01:37 +0000

Asked in a GenAI Engineer interview.

What they are really testing:

Whether you can turn a vague idea of a "good answer" into a repeatable evaluation pipeline that also covers retrieved context and agent memory, not just a single accuracy number.

A strong answer:

Start with a golden dataset of representative queries, reference answers or rubrics, and the expected source chunks or memory items. Include adversarial cases such as stale facts, conflicting memories, and questions that should be refused.
Evaluate each component separately. For retrieval, measure context precision and recall against labeled relevant chunks. For generation, measure faithfulness (claims supported by the supplied context), answer relevance, and correctness against references.
For agents with memory, check what actually entered the context window: was the right memory retrieved, was it recent enough, and did truncation or summarization drop a key fact?
Use LLM-as-judge for scale, but calibrate it against human labels on a sample, pin the judge model and prompt version, and report agreement rates. Prefer deterministic checks (schema validation, exact match, citation presence) wherever they apply.
Run offline evals in CI on every change to prompts, models, or chunking, then score a sample of live traffic online and track drift, user feedback, and escalation rates.
Tie metrics to business outcomes such as task success rate and cost per resolved task, so a higher faithfulness score is not mistaken for a better product.

Common mistakes:

Relying on one aggregate score or on a handful of manual spot checks, which hides regressions in specific query types.
Using the same model as both generator and judge without calibration, which inflates scores through shared biases.
Evaluating only the final answer, so retrieval misses, stale memory, and context overflow get misdiagnosed as model hallucinations.

How would you structure the evaluation for an agent that remembers past sessions, and which failure would you check for first?

Originally shared on LinkedIn.

An AI interview question every day, leadership for tech people and hands-on AI learning:

Opposition didn't stop Indra Nooyi from backing a $3.3 billion bet on Tropicana.

Manohar Papasani — Mon, 09 Nov 2026 04:01:00 +0000

The $3.3 billion Tropicana acquisition, which Nooyi worked on in 1998, initially faced opposition from other PepsiCo executives and Wall Street critics. She had directed the company's global strategy for more than a decade, and she kept pushing the long view.

Her strategy later became "Performance with a Purpose." She sorted PepsiCo's products into "fun for you," "better for you," and "good for you," and moved corporate spending away from junk foods toward healthier options.

Leading people mattered as much as portfolio moves. She wrote to the parents of her leadership team and visited their homes to build a more personal connection, and she worked on a culture of employee retention.

Not every call held. In 2015 she removed aspartame from Diet Pepsi, and in 2016 it was reintroduced after public backlash. She stepped down as CEO on August 6, 2018, after leading PepsiCo since 2006.

“Just realize that when you come to UTA your education is just starting. The world is changing so quickly. Education is a forever commitment. It never stops.”

Indra Nooyi

For engineering and AI leaders, build learning into the plan itself: schedule time for your team to retrain on new tools, and treat a reversed decision as new data, not a failure.

Originally shared on LinkedIn.

An AI interview question every day, leadership for tech people and hands-on AI learning:

Agent pass rates fell from 91% to 15% as codebases grew past 50,000 lines.

Manohar Papasani — Sun, 08 Nov 2026 19:31:00 +0000

The pilot that looks fine on a small repository can tell you little about the large one.

Fast prototypes are not the same as production-ready code.
Here is what the sources say to measure before release.

Does it still work when the codebase or dataset is large?
Are generated data access and permissions reviewed?
Does the AI output need manual refinement?

📌 Scale breaks the pilot
Agent pass rates by codebase size (Sonar)

91% on codebases under 10,000 lines (RepoMod-Bench, 2026, code translation)
67% between 10,000 and 50,000 lines
15% above 50,000 lines

📌 Benchmarks alone do not decide
Pilot before you shortlist (Sonar)

Benchmark rankings alone cannot make the decision, because small isolated task scores rarely predict large codebase behavior.
Shortlist two or three options and pilot them on representative work.

📌 Measure failures, not just wins
Acceptance criteria for AI builders (ToolJet)

Record failures and manual repairs as well as successful generation.
Use the same acceptance criteria and non-production dataset on every shortlisted platform.

📌 Check access before release
Generated data access and writes (ToolJet)

Review generated data access, writes and permissions before release.
Test the current model and builder with the same representative task, recording consumption.

📌 Pricing can shift fast
Microsoft Power Apps

Per-app plan discontinued; entry price jumped from $5 to $20/user/month.
Copilot generates apps from natural language, with over 1,000 pre-built connectors.

What this means for your team:

Pilot two or three platforms on a representative, non-production task.
Compare acceptance rate, rework, cost and quality, not only build speed.
Test on a large codebase or dataset, since small-scale results may not hold.

Pick one real task this week and time how much of the generated output you have to fix.

P.S. Per-user pricing can scale quickly in large teams, so check current pricing before you commit.

Originally shared on LinkedIn.

An AI interview question every day, leadership for tech people and hands-on AI learning:

Seven things happened in AI today; the one worth your attention is OpenAI.

Manohar Papasani — Sun, 08 Nov 2026 12:01:02 +0000

Here is the rundown.

🔬 1. OpenAI: OpenAI announced AI‑generated results that reportedly prove the local smoothing conjecture, which resolves the Kakeya conjecture (announced).
https://champaignmagazine.com/2026/11/01/ai-by-ai-weekly-top-5-october-26-november-1-2026

🚀 2. Hy4: Tencent announced the Hy4 AI model, which it says is trained only on defensive cybersecurity data and excludes material on building hacking tools or harmful biology (announced).
https://www.reuters.com/legal/litigation/tencent-rolls-out-ai-model-trained-with-limited-cybersecurity-data-amid-hacking-2026-11-04

🚀 3. Together Compiler: Together AI announced the Together Compiler, which employs AI agent teams to rewrite GPU code for open‑source models, allowing faster overnight improvements without extra customer effort (announced).
https://www.hpcwire.com/aiwire/2026/11/04/together-ai-launches-together-compiler-for-continual-optimization-of-open-source-ai-models

💰 4. Sentra: Sentra reportedly raised $150 million in a Series C round in November 2026 (reported).
https://aifundingtracker.com/generative-ai-funding-statistics

💰 5. Fal: Fal has reportedly doubled its valuation to $16 billion after a $1.3 billion financing round led by Bessemer Venture Partners (announced). The startup previously raised $350 million at an $8 billion valuation (reported).
https://www.bloomberg.com/news/articles/2026-11-03/ai-inference-startup-fal-raises-capital-at-16-billion-valuation

💰 6. Upwind: Upwind reportedly completed an expanded financing of $400 million in October, as announced by the cloud‑security firm (announced).
https://www.calcalistech.com/ctechnews/article/d1lau0oxq

🚀 7. Images 2.6: OpenAI announced Images 2.6 for ChatGPT, Codex and the API, adding animated looping GIF creation and claiming up to 50% faster generation, with some low‑quality images produced in about six seconds (announced).
https://patmcguinness.substack.com/p/ai-week-in-review-261107

💾 Save this if you are tracking AI releases and funding.
Daily AI digest, with a source link under every item.

Originally shared on LinkedIn.

An AI interview question every day, leadership for tech people and hands-on AI learning:

How do we stop the model from making things up?

Manohar Papasani — Sun, 08 Nov 2026 08:00:57 +0000

Asked in an AI engineer interview.

What they are really testing: whether you treat hallucination as a risk to be managed with layered controls and measured, not as a bug you can fully eliminate. They want to hear about grounding, verification and graceful refusal.

A strong answer:

Start by reframing: LLMs predict likely tokens, so you cannot guarantee factual correctness. Design for bounded, verifiable failure.
Ground answers in retrieved sources (RAG): chunk documents, embed them, run hybrid retrieval (BM25 plus dense vectors), rerank, then pass the top few passages (often 3 to 8) into the prompt with instructions to answer only from them.
Require citations and span-level attribution, and check that each claim maps to a retrieved passage. Return "I don't know" when retrieval confidence or evidence coverage is low.
Lower sampling randomness for factual tasks (temperature near 0) and constrain outputs with a schema where possible, so the model has less room to improvise.
Add a verification layer: an NLI or LLM-as-judge step that checks answer-to-source entailment, plus deterministic checks against tools or databases for numbers, dates and IDs.
Measure it: build a labeled eval set of questions with gold answers, track faithfulness and answer-relevance metrics, and monitor production samples for regressions after each prompt, model or index change.

Common mistakes:

Claiming that fine-tuning on more data will eliminate hallucinations. It can add new facts but does not guarantee the model stops inventing answers.
Relying only on a "do not make things up" system prompt with no retrieval, verification or evaluation behind it.
Ignoring the trade-offs: stricter refusal lowers hallucination but hurts coverage, and extra verification adds latency and cost.

What would you add to this layered approach, or where would you start if you had to ship a grounded assistant next week? Share your answer below.

Originally shared on LinkedIn.

An AI interview question every day, leadership for tech people and hands-on AI learning:

Coach like a river guide

Manohar Papasani — Sun, 08 Nov 2026 04:01:00 +0000

Read the water first: notice what your team is doing today.
Give feedback while the moment is fresh, not months later in a review.
Ask one question before you offer advice or a solution.
Check whether the last course correction really changed the path.
Keep each feedback loop small so people can adjust quickly.
Point out the stretches that flowed well, not only the rapids.
Let people pick their own line through the rapids; you guide them.

A good guide keeps the loop turning: observe, speak, adjust, and repeat.

Originally shared on LinkedIn.

An AI interview question every day, leadership for tech people and hands-on AI learning:

CapCut, Lumen5 and Synthesia each fit a different job, so match the tool to the job before the price.

Manohar Papasani — Sat, 07 Nov 2026 19:30:55 +0000

Lumen5's cheapest paid plan is $19 a month, while CapCut's free tier does not watermark standard exports.

These sources are pricing pages and roundups, not one head-to-head test, so this post compares them without a ranking.

What does it cost to start?
Is it built for editing social clips or for avatar videos?

📌 CapCut
CapCut (ByteDance)

Free tier does not watermark standard exports, per moonb.io
Documents AI video, script and storyboard tools, per ngram.com

📌 Lumen5
Lumen5 plans

Basic is $19.00 per month, for simple unbranded videos, per saasworthy.com
Starter is $59.00 per month and Professional is $149.00 per month; a free plan has limited features

📌 Synthesia
Synthesia

forasoft.com names Synthesia as the pick for avatar videos
Listed in the enterprise layer; these sources give no price

What this means for your team:

For social clips, test CapCut's free tier first.
If you need Lumen5 without its branding, budget from $19 a month on Basic.
For avatar videos, ask Synthesia for pricing before you commit.

Pick the one job your team repeats most this week and test its free option on one real clip.

Originally shared on LinkedIn.

An AI interview question every day, leadership for tech people and hands-on AI learning:

Six things happened in AI today; the one worth your attention is Thinking Machines Lab.

Manohar Papasani — Sat, 07 Nov 2026 12:00:56 +0000

Here is the rundown.

💰 1. Thinking Machines Lab: Thinking Machines Lab is reportedly closing a $5 billion funding round at a $40 billion pre‑money valuation, with Accel leading the round (reported).
https://www.businessinsider.com/thinking-machines-lab-is-closing-5-billion-new-round-2026-11

💰 2. Tutor Intelligence: Tutor Intelligence reportedly raised $75 million in a Series B round led by Andreessen Horowitz, focusing on robot foundation model training and building robot workers (reported).
https://news.crunchbase.com/venture/biggest-startup-rounds-ai-defense-space-tech-shield-zipline

💰 3. Return: Return reportedly secured an €835 million debt financing, which accounted for most of the €1.3 billion total European startup funding tracked between Nov 2‑6, 2026 (reported).
https://tech-insider.org/european-startup-funding-roundup-nov-2-6-2026-60-deals-1-3-billion

💰 4. Alethio Therapeutics: Alethio Therapeutics reportedly closed an oversubscribed Series A of €80.46 million ($90 million), making it the largest round on the week’s list (reported).
https://shattered.io/european-startup-funding-week-hypersonica-100m-nov-2-6-2026

💰 5. Kalepa: Kalepa reportedly raised $75 million in a Series B led by Updata Partners, valuing the AI‑insurance startup at $470 million (reported). The round brings total capital raised to $91 million (reported).
https://www.reuters.com/legal/transactional/ai-powered-insurance-startup-kalepa-valued-470-million-new-funding-round-2026-11-05

🔐 6. Senator Maria Cantwell: Senator Maria Cantwell reportedly released a six‑principle framework urging enforceable NIST safety standards, independent pre‑release audits and mandatory incident reporting for frontier AI companies (reported).
https://www.foley.com/insights/publications/2026/11/blue-wave-new-ai-rules-what-democratic-control-of-congress-means-for-ai-regulation

💾 Save this if you are tracking AI releases and funding.
Daily AI digest, with a source link under every item.

Originally shared on LinkedIn.

An AI interview question every day, leadership for tech people and hands-on AI learning:

Design a RAG system

Manohar Papasani — Sat, 07 Nov 2026 08:00:51 +0000

Asked in an AI engineer interview.

What they are really testing: whether you clarify requirements before drawing components, and whether you know which parts a first version can skip.

A strong answer:

Start with requirements: corpus size and update rate, query types, latency and cost budget, access control (per-document permissions enforced at retrieval time, not left to the prompt), and how quality will be measured.
Ingestion: parse and clean documents, chunk along structural boundaries (often a few hundred tokens with modest overlap), attach metadata such as source, date, version and ACL, and make indexing idempotent so re-runs update only changed documents.
Retrieval: use hybrid search, combining BM25 for exact terms and identifiers with dense embeddings for semantic matches. Apply metadata and permission filters before ranking, pull a broad candidate set (roughly 50 to 100), then rerank with a cross-encoder down to a handful of passages.
Generation: pass the retrieved passages with stable source IDs, instruct the model to answer only from that context with citations, and make "not found in the sources" an explicit, acceptable output rather than a guess.
Evaluation and operations: keep a golden set of questions with expected sources, measure retrieval recall at k separately from answer faithfulness, trace every request, and track latency and cost per query.
Optional pieces: knowledge graphs, multi-agent loops and fine-tuning are not needed for a first version. Add them only when evaluation shows a specific failure they would fix.

Common mistakes:

Drawing the vector database and LLM boxes first without asking about data freshness, scale or permissions.
Assuming the LLM compensates for weak retrieval, and never measuring recall at k.
Applying access control after generation, or relying on the prompt to hide restricted content.

How would you handle a corpus where documents change hourly and users have different access levels? Share your approach below.

Originally shared on LinkedIn.

An AI interview question every day, leadership for tech people and hands-on AI learning:

Hiring for the right thing, not just the right skills.

Manohar Papasani — Sat, 07 Nov 2026 04:00:55 +0000

“Managers are people who do things right and leaders are people who do the right thing. Both roles are crucial, and they differ profoundly. I often observe people in top positions doing the wrong things well.”

Warren Bennis

In engineering and AI, efficiency is easy; direction is hard. A high-performing team isn't built by hiring people who execute fast, but by hiring people who question the goal. If you optimize for speed without validating the problem, you scale failure. Leadership means ensuring the team is solving the right problem before optimizing the solution.

This week, pause your next hiring interview. Ask candidates not just how they build, but how they decide what to build. Look for evidence of strategic judgment, not just technical execution.

Originally shared on LinkedIn.

An AI interview question every day, leadership for tech people and hands-on AI learning: