Aditya Singh — applied AI systems engineer in New Delhi.
I work on the unglamorous half of GenAI: whether an output can be
trusted, traced and shipped. Final year, B.Tech CS at KIIT.

Most of what I build sits around the model rather than inside it —
claim-level verification, evidence linkage, bias auditing, async
workloads that stay upright when the LLM call takes nine seconds.
The interesting failures live in the plumbing.

Python · TypeScript · FastAPI · Next.js · PostgreSQL · Redis · Docker · Prometheus

Previously: Data Analytics @ National University of Singapore · Data Science @ Sukrit Technologies

🛡️ Epistemic Audit Engine — can you trust what the model just told you?

Long-form LLM output gets graded as one blob, or not at all. This splits it into atomic claims, retrieves evidence for each from Wikidata and Wikipedia, verifies them independently, and aggregates the result into a risk score you can act on — so "this paragraph is probably fine" becomes "these two sentences are the problem."

The hard part isn't the checking, it's making the checking reproducible: a deterministic eval harness and an append-only audit log, because a reliability score you can't reproduce is just a vibe.

flowchart LR
    A["Long-form<br/>LLM output"] --> B["Claim<br/>extraction"]
    B --> C["Atomic claims"]
    C --> D["Evidence retrieval<br/>Wikidata · Wikipedia"]
    D --> E["Per-claim<br/>verification"]
    E --> F["Risk<br/>aggregation"]
    F --> G["Scored audit<br/>+ append-only log"]
    E -.->|"no evidence found"| H["Flagged as<br/>unverifiable"]
    H --> F

Loading

FastAPI · Next.js · retrieval · deterministic eval harness   repo →

⚖️ FairHire-AI — resume screening that audits itself for bias

Resume intelligence platform that surfaces hiring bias rather than quietly encoding it. The architectural point: LLM document processing is decoupled from the request path through a worker queue, so a nine-second model call never becomes a nine-second API response.

Instrumented end to end — if a worker is falling behind, the dashboard says so before a user does.

sequenceDiagram
    participant U as Client
    participant A as FastAPI
    participant Q as Redis queue
    participant W as RQ worker
    U->>A: POST /analyze
    A->>Q: enqueue job
    A-->>U: 202 + job id
    Note over A,U: request path never waits on the model
    Q->>W: dequeue
    W->>W: parse · LLM · bias scoring
    W->>Q: store result
    U->>A: GET /jobs/{id}
    A-->>U: result

Loading

FastAPI · RQ workers · PostgreSQL · Redis · Prometheus + Grafana   repo →

🔍 DealLens AI — M&A screening, minus the analyst's weekend

Automates the first pass of investment-banking deal screening: financial ratio filters, NLP over news and filings, synergy detection, and backtesting to check whether the screen would actually have caught the deals that mattered.

Built as a monorepo and hardened for the boring realities — retries, structured logging, health and readiness probes.

Monorepo · Celery · PostgreSQL · observability   repo →

📈 Dynamic Pricing Simulator — what a pricing policy is worth before you ship it

Decision-support system for pricing under demand uncertainty and capacity constraints. Runs candidate policies against simulated demand and benchmarks them on the same footing, which surfaced a 4–5% revenue lift over static pricing in simulation.

Simulation runner · policy benchmarking · Next.js dashboard   repo →

Where the commits are actually landing — the three repositories I touched most recently.

• addyvantage Python · 0★
• TracePack TypeScript · 0★
• smoke-break Python · 0★

Recent public activity

• pushed to addyvantage/addyvantage · 2d ago
• opened issue #44823 in timburgan/timburgan · 2d ago

linkedin.com/in/addyvantage  ·  addy@addyvantage.me  ·  New Delhi, India