Sujit NirmalNow building CallCoach-AILast push 7 Oct 2026

Sujit Nirmal
Role
AI/ML Engineer · Fullstack Architect
Based in
India
Now
AI Engineer, Swayu

I work across machine learning, deep learning, LLMs, retrieval, evaluation, computer vision and product engineering, and I care about what happens once real users arrive: messy inputs, imperfect models, latency, cost, observability, and the engineer who has to operate it.

EmailGitHubLinkedInX

Fig. 1Commit black hole: 1,000 commits across 30 repos, orbiting at Kepler speed. The arc over the top is the far side of the disk, lensed. Click or tap the disk to drop a commit in.
  1. 01Rank: #1 overall of 52 at the WhipScribe Buildathon, Oct 2026 (live leaderboard)
  2. 02Timestamps: CallCoach-AI scores investor calls with four agents, and every flag plays at the second it was said
  3. 03Quotes: FitNova drops any LLM flag that cannot quote the transcript verbatim
  4. 04PyPI package: Sentinel-AI checks numbers in LLM output against the prompt and corrects them

Experience

3 roles since Oct 2024, newest first
Oct 2026 to presentSwayuAI EngineerAgents that research a business, then draft posts, WhatsApp messages, LinkedIn posts and review replies for the owner to approve. I’m building this agent marketing team with the founder.
Nov 2024 to Mar 2025HeadOnAI ResearcherPipelines for text classification, NER and information extraction on HuggingFace Transformers. I coordinated 30+ contributors labelling training data, and ran the precision/recall/F1 evaluations that decided what to fix next.
Oct 2024 to Jan 2025Launchpad.AIAI DeveloperMulti-agent LLM systems (LangGraph, CrewAI) and a Graph-RAG service over Neo4j, shipped to a live customer and Dockerised on GCP. I ran the prompt and retrieval evaluations and the A/B ranking tests, which drove a 2% revenue lift.

Missions

What each one does, what broke, where it runs
M-01

CallCoach-AI

WhipScribe Buildathon winner, Oct 2026, see 05

Four agents score an investor call, and every flag has a ▶ Hear it button that plays the recording from the second it was said.

Anomaly The free-tier analysis server sleeps when idle, so the first score hung. The app now says it's waking the server instead of spinning silently.

LIVESep to Oct 2026Read the mission log

Live call report. 43 flagged moments across a 13:39 call, every quote placed at the second it was said.Live call report. 43 flagged moments across a 13:39 call, every quote placed at the second it was said.
Compliance
Tension
Clarity
Action item
0:003:246:4810:1213:36

12:08 · Compliance · Quote verified in transcript

“By law, you have until midnight of the third calendar day following receipt of the welcome packet to stop the enrollment process.”

Legal rescission right; any obstruction or failure to honor this window is a serious compliance violation.

Fig. 235 flags on the public sample call at callcoachai.sujit.top, 23 verified against the transcript. Hover or tap a marker.

Research

Findings from my own systems
23/35Latest entry · Oct 2026My LLM judge quoted the call. I couldn't find a third of its quotes.Every flag from one scored call, checked against the transcript: where the misses were, what they had in common, and two opposite ways to handle them.Read the note
R-01

Making LLM claims checkable

Ask an LLM for “the line where the advisor overpromised” and it returns a plausible line, said or not. How do you trust a flag?

A flag survives only if its quote is in the transcript, so a reviewer can find every kept flag with a text search.

FitNova write-upR-02

Catching hallucinated numbers

The hallucinations that hurt a business are mostly wrong numbers: refund windows, prices, dates. Can you catch them by checking against the prompt itself?

Yes, when the prompt holds the right value: the wrong number gets caught and corrected before it reaches the user.

Sentinel-AI write-upR-03

Evidence you can hear

A score on a sales or investor call is only useful if the person believes it. What makes an AI judgement believable?

Length predicts a miss. Missed quotes ran 19.8 words against 11.8 for verified ones, and 4 of the 12 stitched two speakers into one quote (Fig. 2).

Field note: what the misses had in commonR-04

Simple models, strong baselines

Datathon 2025: predict item accuracy across stores, under hackathon time limits.

TF-IDF features with logistic regression lifted accuracy from 0.45 to 0.79. I picked a baseline I could explain, then tuned it.

Datathon story

Results

5 results from 20+ hackathons entered
  1. #1 overall of 52WhipScribe Buildathon by Neugence 2026CallCoach-AI67 documented findings against the host platform. The build itself is in 05.
  2. Top 3Datathon 2025 2025Item-accuracy predictionLifted store-level accuracy from 0.45 to 0.79 with TF-IDF and logistic regression. A teammate carried the online round while I was in hospital.
  3. Shortlisted for the finals, Bangalore100x Engineers GenAI Hackathon 2024Dataviz AITurns raw data into animated infographic videos.
  4. Top 30 nationwideHackspiration '24 2024HarvestConnectFarm-to-buyer marketplace, built at VIT Pune Microsoft Learn Student Club's national hackathon.
  5. Shortlisted, internal roundSmart India Hackathon 2024ArogyaKrishiCrop disease prediction and soil analysis for problem statement SIH1638, with team Anant. My most-starred repo.

Mission report

WhipScribe Buildathon by Neugence · Oct 2026
#1overall of 52 builders
  • trackrankscore
  • Track 4 · Invent a workflow#1 / 2475
  • Track 1 · Find what is wrong#2 / 4976
  • Track 0 · Track record#3 / 5569
  • UI challenge 01 · transcript page on a phone#15 / 1929

Judges' note “CallCoach-AI live with a launch film and now one PR with 304 test assertions, the score floor removed; 67 findings, 56 with screenshots, 65 with steps, three reproduced within the hour.”

Open the leaderboard ↗

Fig. 3Standings read from the public leaderboard on 12 Oct 2026. Evaluation is ongoing; ranks use reviewed scores only.

Open source

Status exactly as on GitHub

aden-hive/hive11k★

Ten PRs into an agent framework in January 2026. Nine were closed without merging; the security fix is still open.

4 of 10 PRs shown, status as on GitHub

neugence/whipscribe-buildathon

67 findings filed while building on the platform, 56 with screenshots and 65 with reproduction steps: missing security headers, analytics before cookie consent, API drift from the docs, accessibility gaps.

jundongl/scikit-feature1.6k★

Unit tests for the feature-selection library.

Packages I publish on PyPI

Flight log

7 smaller projects, status as it is on GitHub
ProjectWhat it doesContextStatus★Last push
NeoVerse StoreAR/VR + 3D storefront with an AI conciergeACTIVE2026-09-22
Silent Killer AIAgent that spots workflow inefficienciesLIVE ↗2026-01-06
ArogyaKrishiCrop disease detection and soil analysisSmart India Hackathon 2024VIDEO ↗242026-08-24
Dataviz AIGenAI animated-infographic generator100x Engineers GenAI hackathonVIDEO ↗122025-08-29
Medivance.AIMedical report summarisationVIDEO ↗62025-11-08
Code Quality AgentMulti-language code review with fix suggestionsOFFLINE2026-04-05
ML-Deploy-LiteServes a trained model over REST or gRPCPyPIARCHIVED ↗12024-11-26

Instruments

55 tools across six layers
Fig. 455 tools on three orbits. Lane dots: filled inner, hollow middle, dashed outer. Hover a tool to light its orbit; drag to spin.Tap a tool to light its orbit; swipe to spin.

Field notes

3 notes from my own work, newest first
My LLM judge quoted the call. I couldn't find a third of its quotes.35 flags from one scored call, checked against the transcript, and what the misses had in common2026-10From hospital bed to hackathon: Datathon 2025A teammate carried the online round while I was in hospital; we lifted accuracy from 0.45 to 0.792025-03How I started contributing to open-source ML2024-10All notes, including 7 earlier learning notes
›esc

↑↓ select · ↵ run

Agent view: what an AI agent reads on this site

/llms.txt

Loading…

Tools

    Every mission also has a Markdown twin, e.g. /missions/callcoach.md.