Parsed · document 001

Yash Padam

NAME
Yash Padam
ROLE
AI/ML Engineer
FOCUS
LLMs · RAG · Computer Vision · Vertex AI
AI/ML Engineer. The model extracts. Deterministic code decides.. LLMs · RAG · Computer Vision · Vertex AI
01 / How I build

01 / 06

Something messy comes in.

Real work starts with input nobody designed to be read by a machine.

// scanned handwriting, photos of paper, PDFs of PDFs

02 / 06

A machine reads it.

This part looks like magic. It isn't. It's a guess.

// vision-language model transcription

03 / 06

But it might be guessing.

Ask the same machine the same question three times and you don't always get the same answer.

// N=3 independent runs, self-consistency check

04 / 06

So only what agrees moves forward.

Agreement is the signal. Confidence scores are not.

// per-field consensus; disagreements are flagged, never dropped

05 / 06

Then rules decide — not the machine.

Every decision points at a rule you can read, not an opinion you can't.

// deterministic rule engine, traceable rule codes

06 / 06

And it admits what it doesn't know.

A system that says “I’m not sure” is worth more than one that’s confidently wrong.

// unresolvable fields are surfaced as such, never guessed

BEAT 01 / 06
01 / How I build

How I build

  1. 01 / 06

    Something messy comes in.

    Amox. 500 — 1 tab TDS × 7 days
    ₹ 340 /-   12 / 3 ?

    Real work starts with input nobody designed to be read by a machine.

    // scanned handwriting, photos of paper, PDFs of PDFs

  2. 02 / 06

    A machine reads it.

    medicineAmoxicillin 500 mg
    dose1 tablet · 3× daily
    amount₹ 340
    date12 / 03 ?

    This part looks like magic. It isn't. It's a guess.

    // vision-language model transcription

  3. 03 / 06

    But it might be guessing.

    READ 1
    ₹ 340
    READ 2
    ₹ 340
    READ 3
    ₹ 840

    Ask the same machine the same question three times and you don't always get the same answer.

    // N=3 independent runs, self-consistency check

  4. 04 / 06

    So only what agrees moves forward.

    2 AGREE
    ₹ 340
    FLAGGED
    1 differed

    Agreement is the signal. Confidence scores are not.

    // per-field consensus; disagreements are flagged, never dropped

  5. 05 / 06

    Then rules decide — not the machine.

    does the amount match?yes
    is anything missing?date
    is anything here that shouldn’t be?no

    Every decision points at a rule you can read, not an opinion you can't.

    // deterministic rule engine, traceable rule codes

  6. 06 / 06

    And it admits what it doesn't know.

    RESULTILLUSTRATIVE
    medicineAmoxicillin 500 mg
    amount₹ 340
    datenot determinable

    A system that says “I’m not sure” is worth more than one that’s confidently wrong.

    // unresolvable fields are surfaced as such, never guessed

02 / A small game

Five smudged snippets, the kind my systems are handed every day. Pick what each one says. A machine had three tries at the same ones.

// a toy, not a benchmark — the machine readings are illustrative · keys 1 2 3 · enter

ROUND 01 / 05 · DOSAGEYOU 0 · MACHINE 0
500 mg
SMUDGED
IT SAID500 mg
MACHINE · 3 TRIES500 mg50 mg500 mg
VERDICT
03 / Selected work

Summary

Severity triage, traffic-aware hospital matching, trilingual voice intake. LLM severity assessment with a rule-based fallback when the model is unavailable.

Key mechanics
  • Severity triage
  • Traffic-aware hospital matching
  • Trilingual voice intake
  • Rule-based fallback when the model is unavailable
Stack
Next.jsTypeScriptGoogle MapsRoutesPlaces
github.com/yash-1105/transport-sahayak
Summary

A vision model transcribes; deterministic Python judges.

Key mechanics
  • 18 traceable rule codes
  • N=3 self-consistency consensus
  • Per-field agreement scoring
Stack
FastAPIReactGoogle CloudVertex AI
github.com/yash-1105/rxconcile
Summary

A 4-stage LLM pipeline with prompt caching and streamed stitching.

Key mechanics
  • 4-stage LLM pipeline
  • Prompt caching
  • Streamed stitching
  • .docx / .xlsx export
Summary

Specialised multi-agent pipelines automate POC development end-to-end. 1,000+ passing tests.

Key mechanics
  • Specialised multi-agent pipelines
  • End-to-end POC automation
  • 1,000+ passing tests
The workshop4 more
  • credit-followup-agentrepo ↗

    A finance agent that writes overdue-payment emails, escalating tone across five stages and refusing to send at all once an account needs a human.

    Runs entirely on a local model so client financial data never leaves the machine.

    PythonOllama / MistralPydanticJSON audit trail
  • neurofuzz-traderrepo ↗

    An explainable trading signal system that combines fuzzy logic with neural learning over live market indicators.

    Built so you can see why it made a call, not just what it called.

    PythonInteractive visual analytics
  • Deepfake-minimalrepo ↗

    Lightweight deepfake detection from body pose dynamics rather than faces.

    The idea: fakes get the face right long before they get movement right.

    MediaPipePyTorch (MLP / LSTM)
  • ambitionbox-company-scraperrepo ↗

    An automated collection pipeline that pulls structured company data from unstructured web pages and cleans it into something analysable.

    PythonBrowser automationHTML parsing
04 / Capabilities

AI & ML
  • Generative AI
  • LLM Orchestration
  • RAG
  • LangChain
  • Prompt Engineering
  • PyTorch
  • scikit-learn
  • Transformers
  • Computer Vision (YOLOv5)
  • Speech-to-Text
  • Vector Databases
  • Model Evaluation
Cloud & Platforms
  • Google Cloud
  • Vertex AI
  • Docker
  • GitHub Actions
  • Vercel
  • Render
  • Railway
Backend
  • Python
  • FastAPI
  • Node.js
  • REST APIs
  • Pydantic
  • PostgreSQL
  • SQL
Data
  • Pandas
  • NumPy
  • BeautifulSoup
  • Selenium
  • Feature Engineering
  • Data Cleaning
Frontend
  • React
  • Next.js
  • TypeScript
  • Tailwind CSS
  • Vite
05 / Contact

Yash Padam · AI/ML EngineerD · debug view   / · console