Stephen Thomas, home

Stephen Thomas / AI & ML Engineer

I build AI systems that can show their work.

From messy documents to dependable decisions. I work across retrieval, machine learning, and the software that makes AI useful beyond a demo.

Explore selected workGitHub
Based in
Chennai, India
Works on
retrieval evaluation · document intelligence · backend systems · inference experiments
Diagram: an evidence chainA source page with a paragraph marked A and a table marked B. Each region becomes a ranked chunk: chunk one from paragraph A, chunk two from table B, and a third near miss. The answer at the bottom cites A and B, so every claim can be traced back to a place on the page.FIG. 0EVIDENCE CHAIN · A DIAGRAM, NOT A PRODUCTA · PARAGRAPHB · TABLE12source.pdf · page 12RANKED CHUNKS1chunk 0412paragraph A2chunk 0415table B3chunk 0388near miss4 … 1001 extract02 chunk03 retrieve04 rerank05 citeANSWERevery claim has a pointer[A][B]

01 / Work

Selected work

Three case studies. Each one says what was measured, how, and what the numbers do not show.

A · Retrieval evaluation · research

LocalRAG++

What reranking costs on a CPU, measured, and what its reported gains do and don’t show.

An offline, CPU-only retrieval pipeline and evaluation harness: sparse, dense and hybrid retrieval, cross-encoder reranking, a latency sweep, and a separate study of the document quality that comes before retrieval.

  • BM25
  • FAISS
  • RRF
  • Cross-encoder
  • OCR quality

Co-author, with Dr Salaja S

Read the case studyCode & outputs

Warm mean latency · CPU · reported

6 → 901 ms

BM25, then hybrid retrieval with a cross-encoder reranking 20 candidates. The 95th percentile for c=20 is 1,605 ms.

no reranking+ cross-encoder, c = 10 … 803 ms1 s8 s
Warm mean latency per configuration, log scale. Retrieval scores are in the case study, with their scoring note.
Financial document intelligence: process diagramEight stages from top to bottom: annual report PDF, PDF-first extraction, table-aware chunks, retrieval, structured output, schema checks, reconciliation against a reference workbook, and a dashboard for review. A reference workbook on the right feeds reconciliation and acts as a fallback source for structured output. Values that fail schema checks or reconciliation are flagged down a rail on the left into review.01Annual report PDFlong, table-heavy, layouts vary02PDF-first extractiontext plus table structure03Table-aware chunkstables kept whole, with headers04Retrievalfind where each value lives05Structured outputvalues into a fixed schema06Schema checkstypes · units · required fields07Reconcileagainst a reference workbook08Dashboard & reviewmismatches shown, not hiddenReferenceworkbookExcelfallbackflags → review

B · Extraction · validation · industry work

Financial document intelligence

A figure pulled from deep inside an annual report is only useful if you can say why it’s right.

Pulling specific figures out of long annual reports and making sure they agree with a reference before anyone relies on them: PDF-first extraction, table-aware chunking, schema checks and reconciliation.

Public-safe write-up. No client documents, figures or code; the diagram is a simplified process sketch.

Read the case study

C · SaaS backend · collaborative build

Coleroon.ai

  1. 01OAuth integrations
  2. 02API orchestration
  3. 03Background jobs
  4. 04Quota & cost tracking
  5. 05Publishing workflows
  6. 06Observability

Co-developing a SaaS system where most of the hard parts are other people’s APIs: OAuth integrations, background jobs, quota and cost tracking, publishing workflows and the observability to see them.

Co-developer & technical lead · collaborative build

Read the case study

Index of all work

02 / Approach

How I work

  1. Measure the job, not the model.

    A retrieval system is good if the right evidence turns up, near the top, fast enough to be used. I try to name that job before choosing the clever component.

  2. Keep the boring baseline in the room.

    BM25, a rule, a lookup table. If the simple thing gets within reach of the sophisticated one, that is a finding worth reporting, not hiding.

  3. Write down what a number does not mean.

    Every metric on this site comes with its denominator and its caveats. A result I cannot explain plainly is a result I do not yet trust.

03 / Experience

Where the work happened

Edited down. The long version is on LinkedIn.

  1. Jan 2026 – Present

    ML Engineer Trainee

    Merit Data and Technology

  2. Dec 2025 – Present

    Co-Developer & Technical Lead

    Coleroon.ai

    Collaborative SaaS build: OAuth integrations, background jobs, usage and cost tracking. Case study

  3. May 2025 – Sep 2025

    AI Research Analyst Intern

    BluBridge Technologies

    Rebuilt GPT-2-oriented components; explored Rust/CUDA and custom PyTorch numerical layers.

  4. Feb 2025 – Dec 2025

    ML Engineer

    Rapha Tech

  5. Jul 2024 – Aug 2024

    Intern

    Data Market Cloud Solutions

  6. Apr 2024 – Aug 2024

    AI Developer Intern

    Intel Unnati

    YOLO-based real-time vehicle detection.

  7. Jan 2024 – Mar 2024

    Full Stack Developer

    Edify TechnoSolutions

  8. 2022 – 2026

    B.Tech, Computer Science & Engineering

    Karunya Institute of Technology and Sciences

    CGPA 9.08 / 10

04 / Now

Working notes

Updated November 2026. Small on purpose.

05 / Off the pitch

Outside the editor

Away from the keyboard, I’m usually watching football, playing a little, or overthinking the detail that changed a match. Cricket gets its share of attention too.

Tactics sketchA loose pencil sketch of half a football pitch. A midfielder plays a pass into the half-space, the receiver lays it off, and a runner arrives from deep. The first pass is circled.123the pass that made the chance
The pass before the pass. A sketch, not a scouting report.

06 / Contact

Say hello

Email is the best way to reach me: about roles, research, or anything on this page you think I got wrong.

sthomas.work10@gmail.com