Medaura - Agentic Pharmacy System
Medication errors are an information problem. The information exists - it is just not connected at the moment it matters.
FastAPILangGraphChromaDBLangfuseReact
Pranav DhiranAI Engineer · Researcher
I build systemsthat reason.
I keep digging into rabbit holes, some become systems.
Accepted
Navigate, Don’t Generate
Currently
- 01LFX’26 menteeHyperledger Cello
- 02Neurosymbolic SLMsPre-train + RL
- 03Agri-language modelKnowledge graphs
Everything I build started as a question I couldn't leave alone.
Everything started with a question I couldn't put down. I didn't pick a field so much as a habit — follow a question until it turns into something you can build, test, and hand to someone else. Some questions became models; some became tools; a few became stories worth telling.
Final-year B.Tech, Electronics & Telecom at SGGSIE&T Nanded. Tier-3 college, self-taught in most of what matters.
Recognition
- Qualified — ETHGlobal 2026
- International Finalist (Top 6) — UWA Hack 2026
- National Finalist — Smart India Hackathon 2024 & 2025
- Regional Qualifier — Nxt Wave × OpenAI Buildathon
Every role is a different vantage on the same question.
Apr 2026 – Present
Research Intern
AIISC, University of South Carolina
The bet: a small model with symbolic constraints can do things a large one can't — not in spite of its size, but because of it. I own the pre-training pipeline that tests the bet.
- Own the pre-training pipeline for a 5-person team building an India-focused agriculture SLM — data collection and cleaning through pre-training and evaluation
- Built a 140M-parameter Qwen3-style transformer training/eval pipeline (GQA, RoPE, RMSNorm, SwiGLU, factorized embeddings) over a 6.5 GB / 266K-document corpus, with training-health, domain-validation, and safety evaluation frameworks
- Built the neurosymbolic layer: BPE tokenizer with AGROVOC entity injection, Z3-based causal KG verification, Triple Transformer Encoder
- Built India-Agri-KG — 784 districts, 26 domains, 14,478+ entities, 38,590+ relations — with a 5-stage non-LLM verification layer enforcing source traceability on every triple (unpublished)
- Delivered two technical lectures for the IAIRO-affiliated SLM Bootcamp 2026: Vanilla GPT-2 Architecture, and Scaling Laws & Cost Accounting
Jun 2026 – Present
LFX Mentee
Hyperledger Cello · Linux Foundation
Fabric has an operations problem: steep learning curve, verbose tooling, too much that shouldn't require an expert. The hard part isn't the LLM — it's knowing which API calls compose into what the operator actually meant.
- Building an AI operations copilot for Hyperledger Fabric — natural-language queries to Cello REST APIs through an LLM tool-calling workflow
- Implemented the Django/DRF backend and SSE streaming pipeline powering incremental AI responses in the React dashboard
- Built a Fabric node-logs API on the agent layer with bounded log retrieval and Docker error handling for AI-assisted node debugging
- Developed a multi-party channel invitation workflow end to end: database models, REST APIs, Fabric configuration/signing logic, and dashboard UI
Mar 2026 – Jul 2026
Open Source Contributor
Meshery — CNCF Sandbox Project
The PRs matter less than what you absorb reading other people's production code at scale.
- 5+ merged PRs across Meshery's Go backend and React frontend — service-mesh features, UI components, and API integrations
- Active in issue triage, code reviews, and contributor discussions under CNCF's open-source workflow
Rabbit holes that became systems.
Small Language Model From Scratch - TinyStories
Every LLM course teaches you to call an API. I wanted to know what happens before the API.
PyTorchCustom BPEAMPNLP
03
GNU Radio MCP Server - LLM-to-SDR Bridge
LLMs can reason about RF signals. They just could not touch a radio. This closes that gap.
PythonFastMCPZMQXML-RPCGNU Radio
RF Watch - Open-Source Real-Time RF Spectrum Monitor
Physical-layer first: no protocol decoding, no black-box certainty, just traceable RF evidence.
PythonGNU RadioHackRF OneSignal Processing
The papers that gave me the vocabulary.
Not a reading list — each one changed what I thought was possible.
01arXiv
ReAct: Synergizing Reasoning and Acting
Agents observe before they act.
2210.0362902arXiv
Toolformer: Models Teach Themselves to Use Tools
The mental model for LLMs using external tools.
2302.0476103arXiv
Switch Transformers: Mixture of Experts
Modular capacity beats monolithic scaling.
2101.0396104arXiv
Group Relative Policy Optimization
A foundation for post-training interest.
2402.0330005arXiv
Direct Preference Optimization
Preference optimization without treating RL as magic.
2305.1829001arXiv
ReAct: Synergizing Reasoning and Acting
Agents observe before they act.
2210.0362902arXiv
Toolformer: Models Teach Themselves to Use Tools
The mental model for LLMs using external tools.
2302.0476103arXiv
Switch Transformers: Mixture of Experts
Modular capacity beats monolithic scaling.
2101.0396104arXiv
Group Relative Policy Optimization
A foundation for post-training interest.
2402.0330005arXiv
Direct Preference Optimization
Preference optimization without treating RL as magic.
2305.18290Thinking out loud — same act as building, different medium.
Two sessions taught at the IAIRO SLM++ Bootcamp
PRAMANA Cohort 1 · 2026 — Two lecture sessions taught to the PRAMANA cohort — building GPT-2 from first principles, then what changed and what it cost.
Session 02Vanilla GPT-2 Architecture
A lecture session on the GPT-2 architecture from first principles — how attention, positional encoding, and the decoder stack fit together before any fine-tuning enters the picture.
Watch the session
Session 05Evolution of LLM Design Decisions
A compilation of frontier models case studies — Part 2, Session 05 of PRAMANA: SLM++ Lecture Series. Scaling laws, cost accounting, and how design decisions compound.
Watch the sessionEssays on training, alignment, and going back to first principles
02 publishedIf any of this resonated, let's talk.
Currently interested in AI research internships, open-source collaborations, and systems engineering opportunities. Cold emails work.