Omkar Khadilkar · Applied AI Engineer

Essen, Germany · M.Sc. Big Data & Business Analytics, FOM

I build AI systems that know when they're unsure.

Cheap rules handle the routine cases. A language model only judges the ambiguous rest. Every answer carries a confidence score, a citation, or the word unverifiable.

Seeking a 20-hour/week Werkstudent role in applied AI and data · Essen, NRW or remote · available now · Email me

Proof

2 live demos · 1 public benchmark release · 273 automated tests in 3 repos

Stack

  • Python
  • SQL
  • TypeScript
  • pandas
  • scikit-learn
  • statsmodels
  • dbt Core
  • DuckDB
  • PostgreSQL
  • LLM pipelines (Claude)
  • FinBERT
  • GitHub Actions
  • Next.js
  • Vercel

Languages English C1 (IELTS 7.5) · German A1, learning · Marathi native

02 Flagships

One idea, four domains · every number tagged with how it was measured

01 · Product

Guidance Reconciliation Engine

“Did the company hit the guidance it gave last quarter?”

Decides 181 of 188 forecasts with rules alone. The LLM sees only the unclear 7, and must quote the filing.

  • Python
  • SEC EDGAR
  • Claude Haiku

Live · accuracy pending

Pipeline handles the ambiguous rest
  1. 01 Fetch 8-K releases and XBRL facts
  2. 02 Extract Guidance items
  3. 03 Rules Clean numeric matches
  4. 04 LLM Ambiguous rest, quote-checked

02 · Open source

faultline

“Why did the dbt pipeline break, and where is the evidence?”

Names the exact root cause for 89.6% of 48 dev incidents, with 0 confidently wrong. Held-out score pending.

  • dbt Core
  • sqlglot
  • DuckDB

Benchmark v1 released · triage built · held-out pending

Pipeline handles the ambiguous rest
  1. 01 Gather Monitor anomalies, failed tests
  2. 02 Group Symptoms joined by lineage
  3. 03 Propose Data and code causes
  4. 04 Test Each cause against the data
  5. 05 Rank Survivors, or 'not confirmed'

03 · Product

Werkbank

“Which workshop should take this furniture order?”

Right on 95.8% of the orders it routes on its own. Unsure orders wait for a person.

  • Classifier
  • Postgres
  • PWA · EN/DE

Live

Pipeline handles the ambiguous rest
  1. 01 Request Free text, English or German
  2. 02 Classify Workshop, job type, urgency
  3. 03 Gate Confidence score
  4. 04 Review Unsure orders go to a person

04 · Research

FinBERT fusion

“When should a small model hand a sentence to an LLM?”

Reaches 0.878 accuracy by sending only the unsure 12% to an LLM. The gain is 5 sentences, reported as such.

  • FinBERT
  • Confidence gating
  • SEC filings

Research · code being published

Pipeline handles the ambiguous rest
  1. 01 Filings Risk Factors and MD&A, 6 firms
  2. 02 FinBERT Fine-tuned, 3 classes
  3. 03 Gate Confidence below 0.75
  4. 04 LLM Claude Sonnet fallback

03 Index

All projects →
No.ProjectTypeWhat it showsLink
05 Sustainability commitment classifier Tool Only 12.3% of 1,562 climate commitments come with a number ↗
06 DAX 40 capital structure Research More profitable DAX firms carry less debt · graded 1.3
07 ESG and bond spreads Research An honest null: ESG rating changes did not move bond spreads
08 Dem Fragen Product Plain-language answers under Indian law, with sources · real eval pending ↗
09 GradeLens Tool Shows FOM students what each exam needs for a target grade ↗
10 profile-hud Tool GitHub profile cards built from real repo data · 156+ tests ↗

04 Currently

Studying

M.Sc. Big Data & Business Analytics

FOM University of Applied Sciences, Essen

Building

faultline

An open-source reliability assistant for dbt pipelines, with its own fault-injection benchmark

Available

20 hours a week, from now

Werkstudent roles in applied AI and data · Essen, NRW or remote

Email me