Ghanshyam Paunikar

LLM Engineer, Sigma AI

LLM Engineer at Sigma AI, London, UK, from April 2024 to September 2026 (about two and a half years).

• Built LLM-powered applications end to end, from backend services and model orchestration to the review tools used daily by 300+ transcribers and reviewers.
• Developed LLM evaluation frameworks with golden datasets scoring accuracy, groundedness, faithfulness and hallucinations, plus automated regression benchmarking of prompt chains before every release.
• Defined and monitored LLM performance metrics with Langfuse and OpenTelemetry, and ran experiments comparing models and prompts, including fine-tuning with LoRA/PEFT.
• Built LLM workflows and agent-style chains with LangChain and LangGraph using multi-step reasoning, tool use, and retrieval for grounding.
• Built data pipelines and Python APIs (FastAPI) integrated with internal data sources, with logging, monitoring and data privacy controls.
• Fine-tuned Large Language Models with LoRA/PEFT on transcription and annotation datasets, improving guideline adherence and output consistency.
• Built LLM-powered QA pipelines that detect transcription errors and formatting deviations automatically, cutting manual review effort and accelerating delivery timelines.
• Built evaluation and regression harnesses with golden datasets scoring groundedness, faithfulness and accuracy, run before every release to catch errors early.
• Added guardrails and Pydantic schema validation to reduce hallucinations and formatting failures, and instrumented pipelines with Langfuse and OpenTelemetry for cost, latency and quality.
• Engineered automated data processing and evaluation pipelines in Python, decreasing analysis turnaround time by 30%.
• Designed real-time dashboards tracking annotator productivity, agreement scores and quality metrics, giving leadership actionable operational insights.
• Worked with platform, data and operations teams to ship production-ready AI, reducing error rates by 12%.

Technologies: Python, LangChain, LangGraph, LLM APIs, LoRA/PEFT, Hugging Face Transformers, PyTorch, Pydantic, pytest, FastAPI, Docker, Langfuse, OpenTelemetry, SQL

Data Scientist Intern, Virtual Engineering Centre

Data Scientist Intern at the Virtual Engineering Centre, University of Liverpool, Liverpool, UK, from October 2023 to March 2024.

• Built an LLM integration for automated case-study processing in Python: Google Forms collected structured submissions, OpenAI API calls handled prompt-based processing, and outputs were generated automatically, cutting manual processing time by about 60%.
• Embedded generative AI in an existing business workflow used for real business case-study submissions, from input collection through API calls to automated handling of each case.
• Evaluated results against the manual process with technical stakeholders and reported performance through Power BI dashboards for project progress and resource allocation.
• Developed Python-based automation in NVIDIA Omniverse, reducing repetitive 3D layout design effort by about 40% during project testing.

Technologies: Python, OpenAI API, SQL, TensorFlow, Tableau, Power BI

Data Scientist, Zummit Infolabs

Data Scientist at Zummit Infolabs, Bangalore, India, from February 2021 to October 2021.

• Architected and fine-tuned a UNet-based deep learning model for MRI image segmentation using PyTorch, achieving a 15% improvement in cancer detection accuracy.
• Built end-to-end data preprocessing and model deployment pipelines with augmentation and normalisation, reducing training time by 20%.
• Benchmarked state-of-the-art computer vision architectures against baseline models, translating findings into design decisions that improved segmentation performance.
• Communicated complex deep learning findings to non-technical medical stakeholders weekly, bridging ML research and clinical decision-making.

Technologies: Python, PyTorch, OpenCV, NumPy, Pandas, Scikit-learn, Matplotlib

AI NPC: Living Characters

AI NPC: Living Characters — GPT-3.5, Unity, C#, Meta Voice SDK (Wit.ai), Oculus LipSync. His MSc dissertation project. NPCs you talk to by voice: Wit.ai speech-to-text, GPT-3.5 in character with the conversation history, Wit.ai text-to-speech, and lip-synced facial animation; the model can trigger in-game actions by keyword, an early form of tool calling. Reported a 45% engagement lift. Repository: github.com/GhanshyamPaunikar/AI-NPC

Innerbloom (AI journal)

Innerbloom — Ollama, FastAPI, local JSON storage. A 100% local AI journal with zero telemetry. Features streaming agent chat with three personas (Companion, Observer, Challenger), 16-emotion tagging, a Wellbeing Radar tracking burnout and spiral patterns, a People social CRM, and a Memory Graph. Everything runs on the user's own machine via local LLMs and semantic search; entries are analysed for emotion, intensity and themes, and the journal can be queried conversationally to surface contradictions, emotional triggers and long-term patterns. Repository: github.com/GhanshyamPaunikar/innerbloom

CarbonWatcher

CarbonWatcher — Python, Intel RAPL, NVIDIA NVML. Installable via "pip install carbonwatcher". Tracks energy consumption, CO2 and cost for ML training runs and LLM API calls across Anthropic, OpenAI, Ollama and Hugging Face. Includes location-aware grid intensity, a CLI with monthly dashboards and team reports, and 44 pytest tests. Repository: github.com/GhanshyamPaunikar/carbonwatcher

F1 Race Predictor

F1 Race Predictor — XGBoost, HistGBR, LightGBM, FastAPI, Vue.js, SHAP; was deployed on Railway. An ensemble (including a LightGBM learning-to-rank model) trained across 8 F1 seasons (about 3,500 driver-race rows, 22 leakage-safe features including weather and practice pace). Runs 2,000 Monte Carlo races for win probabilities, predicts DNF risk, and explains each driver with SHAP. An earlier 12-feature version scored a cross-validated MAE of 3.41 positions. Podium and full 20-driver grid predictions. Repository: github.com/GhanshyamPaunikar/f1-race-predictor

Image Resolution Upscaler

Image Resolution Upscaler — Stable Diffusion x4, PyTorch, diffusers, Gradio. Prompt-guided 4x super-resolution via latent diffusion, hallucinating sharp high-frequency detail at four times the original resolution. Configurable diffusion steps and noise levels; auto-caches a 1.7 GB model on first run. Runs locally with a Gradio UI. Repository: github.com/GhanshyamPaunikar/Image-Resolution-Upscaler

ASL Sign Language Recogniser

Sign Language Recogniser — MediaPipe, PyTorch, OpenCV. Real-time ASL alphabet reading from 21 MediaPipe hand landmarks using geometric rules on finger positions (about 10 letters out of the box), then word-level work: a PyTorch Transformer scaffold for ASL words and a BSL recogniser on his own scraped dataset (741 clips, 100 words) with a Bi-LSTM, evaluated on held-out signers (16-18% top-1 on 100 words), plus text-to-sign and a browser demo. Repository: github.com/GhanshyamPaunikar/Translator-Sign-Language

SpaceX Falcon 9 Landing Prediction

SpaceX Falcon 9 Landing Prediction — his IBM Data Science Capstone. Predicts whether the Falcon 9 first stage will land successfully, enabling launch cost estimation. Built with classification models, exploratory data analysis, SQL, Folium maps and a Plotly Dash dashboard. Repository: github.com/GhanshyamPaunikar/IBM-Capstone

Zeul (Android wallpaper app)

Zeul — Kotlin, Jetpack Compose, Material 3, Supabase (Postgres). A free, open-source (GPL-3.0) Android wallpaper app: curated HD and AMOLED feeds, a "For You" feed that learns from picked styles, search with tag autocomplete and colour/device filters, full-screen preview with fake home and lock screens, apply with crop, blur and dim, community uploads, favourites that work offline, and an auto wallpaper changer. Launched 12 October 2026. Built by two developers in about two weeks, each working with their own Claude Code agent and coordinating through a written contract of shared tables, RPCs and interfaces, a task board and hand-off notes in git, with CI building and testing every push. Ghanshyam built the Android app: UI, data layer, preview, search, onboarding, tablet layouts and background work; his co-developer Raunak Patil built the backend, dashboard, ads and billing, design system and launch. Repository (Raunak's account, both credited): github.com/raunakpatil/ZeulWallpaperApp

Piano Tutor

Piano Tutor — JavaScript (plain ES modules, no build step), HTML, CSS, Electron for the Mac and Windows apps, Web MIDI, with Python tooling. A Synthesia-style app for learning piano: notes fall toward a keyboard where every key is named (C D E or Do Re Mi). Listen, Practice (the song waits for the right note) and Play along modes; left hand, right hand or both; tempo down to 50%; loop any section; a scrolling grand-staff sheet music view generated from the song; metronome; keyboard sizes from 25 to 88 keys with out-of-range notes folded into range; sustain pedal; 30-second previews; a MIDI library from any folder; works offline. Works with any USB or Bluetooth MIDI keyboard, or the computer keyboard. Free, with a one-time purchase to remove ads. Released October 2026. Repository: github.com/GhanshyamPaunikar/piano-tutor

MSc, University of Liverpool

MSc in Data Science and Artificial Intelligence, University of Liverpool, UK, September 2022 to September 2023. His MSc dissertation was the AI NPC: Living Characters project — a modular NPC system with AI-driven dialogue, real-time voice recognition, lip-sync animation and sentiment-aware emotional responses.

B.E., Raisoni College

B.E. in Mechanical Engineering, Raisoni College, RTMNU (Rashtrasant Tukadoji Maharaj Nagpur University), Nagpur, India, 2016 to 2020. An engineering degree, which he followed with an MSc in Data Science and AI.