Aksh Goyal

Aksh Goyal

I build AI systems that have to be right — and the tooling that proves whether they are.

Fifteen months embedded with the Vaccines Commercial team of a global pharmaceutical company, building the RAG systems and internal tools their marketing, medical and bidding teams use. IIT Delhi, 2025.

Open toAI EngineerForward Deployed EngineerAI Product
See what I've builtProduct decksResumeEmail me

Experience

Analyst — Aspect Ratio

Jun 2025 – Aug 2026

AI & strategy consulting. For 15 months I was embedded with the Vaccines Commercial team of a global pharmaceutical company — sitting with their marketing, medical and global bidding teams, turning briefs with no internal precedent into software they use.

94%
accuracy reached by the chatbot, once I had an evaluation framework measuring it
40+
countries the vaccine tender tool covers, across 5 vaccines
10
users who validated the chatbot before it was handed over
  • Owned an HCP persona-simulation chatbot end to end, from a brief with no internal precedent — framed the problem with the client’s marketing team, designed the architecture, shipped it over a local network. Validated by 10 users, then handed to a large consultancy for production build-out.
  • Architected its multi-agent RAG backend: routing agents directing queries across a hybrid retrieval stack (semantic vector search + BM25).
  • Built the evaluation framework that took it to 94% accuracy — designed test scenarios and failure paths, then iterated chunking, embedding and search strategy against measured results.
  • Built a tender guidance tool for the global vaccines bidding team spanning 40+ countries and 5 vaccines, putting pricing and bid-strategy guidance at the point of decision.
  • Wrote 4 web scrapers turning undocumented conference sites with no API into structured datasets, and ran an end-to-end SQL analysis of vaccine-effectiveness data behind a European surveillance report.

Read how the chatbot was built

Summer Trainee — Grant Thornton Bharat

2023 & 2024, public sector consulting

Two summers on government engagements: a Power BI dashboard tracking a large-scale resource deployment state-wise and designation-wise, and — the year before — unifying citizen datasets in Alteryx behind a Tableau dashboard covering 15+ schemes, reviewed quarterly by state leadership.

More background

How I build

Five things I do consistently. Each one points at something specific in the work below, not a general belief.

Measure first, then tune
RAG systems are easy to fiddle with and impossible to improve blind. On the chatbot I wrote the evaluation scenarios before touching chunking or search — which is the only reason 94% means anything. The eval harness takes that further: it scores retrieval and generation separately, so a wrong answer says which half to fix.
Ordinary code by default, the model only where judgment is needed
The outreach agent calls a model once per contact, to write. Whether an email is allowed to go — grounding, dedupe, opt-outs, the daily cap — is decided by plain Python that can be tested without a network.
Constrain the output, then verify it anyway
The drafting schema only accepts a contact’s real field names, so a model cannot cite a source that doesn’t exist. Then code checks every name, place and number against that contact’s data. Plan My Day does the same with task ids: one that matches no real row never reaches the screen.
One boundary for every model call
In the eval harness, every call passes through one module that caches on the full request, enforces a hard spending limit, counts retries and pins the provider. In the outreach agent, a test fails the build if anything outside the adapters imports a vendor library.
Write down the limits
The outreach checker has one known blind spot, and it is recorded as a test so it can’t be forgotten. The eval harness counts errors next to scores, so 90% never quietly means 90% of the cases that didn’t crash. Plan My Day has no auth, so it binds to 127.0.0.1 and says so.

Projects

Built outside work. Each page covers the problem, the architecture, and the decisions that were actually hard — including the ones that went wrong first.

  • RAG Eval Harness

    2026 · in progress · Python · NumPy · Pydantic · OpenRouter

    Measures whether a change to a RAG system actually helped — and when it didn’t, says whether retrieval or generation broke.

    Read the case study

  • Outreach Agent

    2026 · Python · SQLite · Pydantic · SMTP / IMAP

    Personalised B2B email that can’t invent a fact about the person it’s written to — every claim is checked by code against their own data before a human approves it.

    Read the case study

  • Plan My Day

    2026 · Python · FastAPI · SQLite · Gemini

    A productivity coach that breaks goals into tasks you can actually start, and tells you which ones to do today.

    Read the case studySource on GitHubRun it yourself

Also built

GEO optimization engine (Jul 2025) — rewrites webpages so that language models retrieve and cite them. Evaluated against 150+ synthetic LLM queries; citation rate improved 41%. Not open source, but the evaluation-first pattern is the same one I use everywhere else.

Product Thinking & Problem Solving

Strategic case studies and presentation decks tackling complex systems from first principles — focusing on root cause analysis, operational trade-offs, phased roadmaps, and measurable metrics.

Contact

Email is the fastest way to reach me. Based in India, open to relocating.

Email
aksh.iitdelhi@gmail.com
GitHub
github.com/akshgoyal
LinkedIn
linkedin.com/in/akshgoyal