hello, I’m
AditiSoni
AI engineer
I have a little chaotic clutter in my head, insane ideas and even wild aspiration. But atleast nobody can say that I am without a plan ;)
a little about me
Before we get into any serious details ~
"Museum of failures or gallery of trying?" I have an unholy belief in the fact that if I set my mind to something, it's getting completed. Completely toxic self-drive but it motivates me so we keep at it.
Most of my work sits where machine learning meets rigour figuring out whether a model did the right thing, and building systems that can say so reliably. I listen to Lana Del ray, Tame Impala, NBHD, and (who m I kidding?) all pop girls hehe.
I am also basically a nerd coz u will either find me buried beneath a good Novel or manhwa. Need good recs? I ur gal. I am also bullheaded so "I can't do it" isn't in my dictionary.
Currently: Caffeinated
Probably: Reading a research paper
what I work with
Skills & technologies
Machine learning
PyTorch, TensorFlow, EfficientNet-B0, WavLM, DistilRoBERTa, Whisper, SCRFD, scikit-learn, spaCy
Evaluation & agents
Agent harnesses, VLM-based judging (Gemini), execution-trace analysis, automated validation and sanitisation
Vision & data
OpenCV, albumentations, NumPy, Pandas
Backend
FastAPI, Django REST, Celery, Redis, PostgreSQL, Gunicorn
Scraping
Scrapy, Selenium, sitemap parsing, DFS crawling
Delivery & formal methods
Docker Compose, GitHub Actions, CI/CD, AWS S3, Lean, mathlib
things I’ve built
Projects
OmniSent
A multimodal sentiment engine that reads faces, voices and text together and turns emotion into UX scores, flagging events at exact timestamps.
- Confidence-weighted late fusion with graceful degradation when a signal drops out.
- SCRFD face detection into a fine-tuned EfficientNet-B0, trained on 70,000+ images.
- Fine-tuned WavLM for voice tone; Whisper into a fine-tuned DistilRoBERTa for text.
Smart Resume Analyzer & Matcher
Ranks candidates against a job description, so the best fits surface first instead of getting buried in a pile of PDFs.
- spaCy extracts skills and experience; TF-IDF and cosine similarity score each match.
- Async Celery workers on Redis parse PDFs, so uploads never block the API.
- Containerised with Docker Compose and deployed through CI/CD.
where I’ve been
Experience & education
AI Engineer · Parsewave
- Designed and scaled an evaluation task pipeline: a custom agent harness generates video-editing tasks and validates them before they reach a benchmark, enforcing oracle correctness and NOP constraints.
- Built multimodal evaluation workflows with VLM-based judging (Gemini 3.5 Flash) for reproducible benchmarking.
- Cut defective submissions by 80% by analysing agent execution traces and adding automated validation.
Data and Tech Intern · GrowthforImpact
- Saved 50+ hours/week with a scalable scraper over 5,000+ company domains.
- Improved candidate-matching accuracy with a multi-stage normalisation pipeline.
- Automated scraping with GitHub Actions to keep data fresh for continuous MVP deployment.
B.Tech, Computer Science & Engineering
- Organiser, Stellaris 2026: a 12-hour hackathon run by GeeksforGeeks x Elixir Tech Community at ABES Engineering College, with 1200+ registrations.
- Four tracks (AI/ML, Web3, Fintech and Open Innovation), with Major League Hacking as global partner and Devfolio as platform partner.
- Brought in sponsors including GitHub and Risein (gold), Cyrene AI and CertifyX (silver), and a bronze tier with Replit, Windsurf and Stellar, plus community partners from colleges across the region.
- Prize pool of $550 across three places, with mentorship, certificates and goodies for the winners.
come say hi
Let’s build something careful and kind.
If you’re working on AI evaluation, multimodal systems or applied ML and need someone who checks their work, I’d love to hear from you.
