Adarsh S Mcold start
000
resolving runtime01/05
← All work

Fine-tuning & RAG

Production LLM work across finance and developer tools

Two and a half years of shipping adaptation and retrieval systems for paying clients: fine-tuned LLMs for code generation and intent recognition, and RAG pipelines over financial documents.

Where
Fiverr — freelance
When
2024 — present
Role
Machine Learning Engineer & Gen AI Specialist
Stack
PyTorch · GPT · BERT · FAISS · Pinecone · Transfer learning · Prompt engineering
30%Contextual accuracygain from fine-tuning GPT / BERT
40%Manual effortreduction on financial data processing

Problem

Client work has a constraint research does not: someone is paying, and they will not accept a demo. Both engagements here started with a general-purpose model that was directionally right and practically unusable — a code assistant that misread intent, and a financial workflow where a human still re-read every document to find four numbers.

Constraints

  • 01Fixed freelance budgets — no room for a training run that might not work
  • 02Client data, client infrastructure, client compliance rules
  • 03Handover matters: the client has to be able to run it after the engagement ends
  • 04Measurable against the manual process it replaces, not against a benchmark

Architecture

SOURCEINDEXRETRIEVEADAPTDELIVEREVIDENCEFinancial documentsclient corpusCode + intent datatask corpusEmbeddingschunkedFAISSlocal vector indexPineconemanaged indexRetrievaltop-k evidenceFine-tuned GPT / BERTtransfer learningPrompt engineeringtask framingClient workflowhandover-ready
fig.01Two pipelines that share a shape: adapt the model to the domain, retrieve the evidence, and keep a human where the cost of being wrong is high.(scroll horizontally on narrow screens)

Key decisions & trade-offs

  • Transfer learning before anything more ambitious

    Fine-tuning from a strong base with transfer learning is the highest-certainty use of a fixed freelance budget. It gets to a measurable improvement fast, and it leaves room in the budget to iterate if the first attempt lands short.

    Trade-offIt caps the ceiling — you inherit whatever the base model is bad at. For a client who needs a working system this quarter, a reliable 30% beats a speculative 60%.

  • FAISS and Pinecone, chosen per engagement

    Where data could not leave the client's environment, FAISS ran locally. Where the client wanted no infrastructure to operate after handover, Pinecone was the honest answer. The retrieval logic stayed the same either way.

    Trade-offSupporting both means an abstraction layer and two sets of operational quirks. It is the price of not forcing every client into the same deployment story.

  • Measure against the manual process, not a benchmark

    The number a client cares about is how much work disappeared. Framing the financial pipeline as 40% less manual effort — rather than a retrieval score — is what made the value legible to the person approving the invoice.

    Trade-offIt makes results harder to compare across engagements, and it means baselining the human process first. It is also the only number anyone acts on.

Results

30%Contextual accuracycode generation and intent recognition
40%Manual effortfinancial data processing
87%Churn model accuracyScikit-Learn + XGBoost
92%CNN classificationimage classification via transfer learning
2 yrs 7 mosEngagementongoing, repeat clients