Available for opportunities · Bangalore, India · Remote

HarshaaVardhana

Foundational AI Researcher (NLP / GenAI)

Agentic AI researcher and applied mathematician, building large action models, agentic workflows and AI data analytics tools, and publishing in NLP and deep RL. Stochastic, never random.

Illustrated portrait of Harshaa Vardhana KV
  • JEE Main7,7862023
  • KCET1,2582023
  • VITEEE2,5042023
Show

Work

AI Full Stack Engineering Intern, DocNexus.ai

A US healthcare claims-analytics company: forecasting and estimation systems, and the AI agents that answer questions over the same data.

  1. FinSignalForecasting a drug’s quarterly revenue from claims before the company reports it, then letting the SEC score the forecast.ShippedJun 2026 – presentAI / ML, Quant & Math, Data & Analytics9.42%median error, October model
  2. Full-history claims cost estimatorCosting every medical claim line in a decade of US claims, and discovering that half of them never needed estimating.ShippedJun – Aug 2026Data & Analytics, Quant & Math−44.4%line bias before
  3. In/out-of-network from payer adjudicationStop inferring the network. Read what the payer said about it.ShippedJul – Aug 2026Data & Analytics40.67%unknown lines recovered
  4. Front-door query routerDeciding, before any LLM runs, whether a question needs a lookup or an agent.ShippedJul 2026AI / ML, Full stack97.5%held-out accuracy
Show all DocNexus work (11 more)
  1. PayerLensA Medicare Advantage panel and leakage worklist that says “here is our reasoning, check it” instead of “trust our score”.ShippedJul – Aug 2026Data & Analytics, Full stack
  2. Market sizing engineNational market size and share for radiopharmaceuticals, from a partial claims panel, with every number carrying its source and its band.ShippedJul – Oct 2026Data & Analytics, Quant & Math, Full stack49products sized
  3. Answer-consistency audit → knowledge mapAsked the same question five times, the agent gave five different answers. Finding out why.DesignAug – Oct 2026AI / ML, Data & Analytics, Research
  4. Self-verifying answers on any warehouseAttaching a verified / unverified verdict to every aggregate answer, without gold labels.In progressOct 2026 – presentAI / ML, Research, Data & Analytics
  5. Query-intent landscapeWhat do people actually ask a clinical search engine? Three methods, one vocabulary.ShippedJul 2026AI / ML, Data & Analytics
  6. Drug-to-security crosswalkMapping every drug claim since 2017 to the company that sold it on that date, and to its listed security.In progressOct 2026 – presentData & Analytics, Quant & Math

Smaller pieces

  • Org-private wiki agent

    Heading-aware chunking, embeddings in pgvector with an HNSW index, and a ReAct agent that answers with citations and filters every query by organisation. Built as a new agent rather than reusing the file agent, because the file agent silently truncated long documents.

    Shipped
  • HCP mobile app features

    Live search on the real index, PDF export through the share sheet, and external booking hand-offs. Found that free-text specialties returned nothing because the index needs exact taxonomy strings, so specialty chips now map to groups of verified values. Raised chip contrast to 16.5:1 and cut an accidental lockfile churn down to the real change.

    Shipped
  • Medical-code resolver audit

    A line-by-line comparison of two natural-language → billing-code resolvers, with live index probes. Found three failure classes (an ungrounded guess, a source too small to ground against, and a real code that is the wrong code) and designed a single grounding core: live indexes for categorical fields, curated version-locked value sets plus invariant checks for clinical codes.

    Design
  • Hybrid drug mapping

    Hardened a drug-name mapping layer that routes deterministic matches first and sends only the remainder to GPT-4o-mini, grounded in RxNorm, alongside a dashboard-template library.

    Shipped
  • Compliance-gated conversational rep

    Contributed to the design of an AI rep that talks to physicians: grounded-only answers, verbatim safety text, adverse-event routing to pharmacovigilance, prompt-injection blocking and fail-safe degradation, with per-programme storage isolation enforced by construction.

    Contributed

Earlier experience

  1. Jan 2026 – Mar 2026Foundational AI Researcher, NLP & Generative AIOpen Source Program Initiative (OSPI '26), IEEE Computer SocietyResearch on knowledge conflicts in retrieval-augmented generation, mentored by PES alumni at American Express and Cisco. Co-authoring Axiom; presented the proof of concept at HiDevs, Bangalore.
    • Training-free activation steering that pushes a model to follow retrieved context when it contradicts what the model memorised, with per-query layer selection (BCILS) and a budgeted subspace search (B-PLIS, CMA-ES).
  2. Jun 2025 – Aug 2025Research InternCentre for Robotics, Automation and Intelligent Systems (cRAIS), PES UniversityDeep reinforcement learning for mobile-robot navigation in unknown environments, published at IEEE ICSET 2025.
    • Compared DDPG warm-started from a GRU trained on A* demonstrations (with prioritised replay) against SAC with LSTM actor and critics, in a LIDAR simulation.
    • Designed reward shaping that balances goal progress, heading, proximity penalties and a step cost, and checked convergence across maps and obstacle layouts.
  3. Oct 2024 – Aug 2025Cybersecurity InternSecurity Operations Centre, Center for Information Security, Forensics and Cyber Resilience, PES UniversityMalware analysis and network-defence work in the university security operations centre.
    • Completed TCM Security’s Malware Analysis course and analysed the Mydoom worm’s behaviour, propagation and code structure.
    • Simulated volumetric DDoS traffic in an isolated, authorized lab to validate detection and mitigation controls, using nmap and Wireshark for packet-level analysis.
    • Contributed to incident-response playbooks for DDoS triage, suspicious binary execution and first-response forensics.
  4. Feb 2025 – Apr 2025SkyScan: drone for 3D pollution mappingCenter for IoT, PES UniversityA sensor-equipped drone that maps air quality, temperature and humidity across altitudes, addressing the blind spots of static monitoring stations, on a YMFC-AL flight-controller stack.
  5. Jun 2024 – Jul 2024AI Research Intern, AutodroidPESU Venture Labs, Deep Tech InternshipA large-action-model pipeline on Android that maps instructions to UI actions, with error recovery when layouts change and guardrails on agent actions. The project was discontinued; it shaped the later agent work.
  6. Feb 2024 – May 2024Web Development InternAatmatrisha, PES UniversityBuilt schedule, event and registration pages for the university’s annual techno-cultural festival website.

Projects

Selected projects

Archive

Research

Papers and ongoing research

  1. Context-Aware Hierarchical Summarization of Long Documents via Adaptive Chunking and Semantic Boundary DetectionSubmittedEMNLP 2026, BudapestOct 2026

    Harshaa Vardhana KV, Himani Pranesh Rao, Harshith P, Gowri Srinivasa

    Mentor: Dr. Gowri Srinivasa

    Long documents overflow a summariser’s context window. The pipeline splits at semantic boundaries found with sentence embeddings instead of fixed windows, sizes chunks by document length, summarises each with BART and merges recursively.

    My part. Built the evaluation and benchmarking pipeline: the ROUGE and BERTScore comparisons, and the precision–recall and efficiency analysis.

    • ROUGE-1 of 0.3386 on 150 long BookSum documents, against 0.2800 for fixed chunking and 0.1481 for truncation (+21.0% over fixed chunking), with ROUGE-1 recall of 0.5557.
    • Preferred over a gemini-2.5-flash summary on 15 of 19 documents in a head-to-head subset.

    Limits. Recall gains trade off against precision. The baselines ran with a smaller final summary budget than the proposed method, which favours it on recall. One domain and one base model.

  2. A Deterministic Multimodal Earnings Call Analysis Pipeline with FinBERT Embeddings and Rule-Based Transcript and Presentation ExtractionAccepted12th IntelliSys 2026, AmsterdamSep 2026

    Harshith P, Godha Kanchi Mandya, Harshaa Vardhana KV, Ashwini M Joshi

    Mentor: Dr. Ashwini M Joshi

    Predicts the direction of a stock’s move after an earnings call by fusing FinBERT embeddings of transcript sections with structured financial features and features extracted from the presentation deck, in an inspectable, deterministic pipeline.

    • Text alone is weak (macro-F1 0.207). Multimodal fusion with a calibrated soft ensemble reaches 58.68% accuracy and macro-F1 0.587.

    Limits. FinBERT is frozen, not fine-tuned. The rule-based extractors are brittle on unusual document formats.

  3. DRL-based Navigation of Mobile Robot in Unknown EnvironmentPublished2025 IEEE 15th International Conference on System Engineering & Technology (ICSET), Kuala LumpurDec 2025

    Daniya Rehman, Harshaa Vardhana KV, Nagarjun N H, Karthikeya Katragadda, Rashmi Ugarakhod

    Mentor: Dr. Rashmi N Ugarakhod

    Warm-starts deep RL agents with sequence models trained on A* demonstrations, and compares DDPG with a GRU warm start against SAC with LSTM actor and critics for LIDAR-based navigation.

    • SAC with LSTM reached a cumulative reward of 1513.36, against 1376.84 for DDPG with the GRU warm start.
  4. Axiom: Resolving LLM Knowledge Conflicts in RAG via Bayesian Causal Activation SteeringIn progressOSPI '26, IEEE Computer Society2026

    Deeptanshu Kumar, Harshaa Vardhana KV, S Mayur, Shailesh

    Retrieval-augmented models often ignore the retrieved context in favour of what they memorised. Axiom is a training-free intervention: steering vectors applied to internal activations, with a Bayesian per-query layer selector (BCILS) and a budgeted subspace search (B-PLIS), triggered only when a conflict detector expects one.

    • On ConFiQA with Gemma 3 4B (1,500 examples per subset), steering at the best static layer raises context-faithfulness from 0.755 to 0.817.
    • BCILS matches the best static layer (0.816) rather than beating it: it selected that same layer for every query. Reported as a match, not a win.

    Limits. The conflicts are counterfactual, not natural. Answers are matched by substring. Two models tested.

  5. ORACARE: A Multimodal Digital Twin of the Oral Cavity for Predictive Diagnostics and Treatment Simulation of Oral CancerIn progressCapstone, PES UniversityMay 2025 – present

    Mentor: Dr. Jayashree R

    A patient-specific virtual twin for simulating oral-cancer growth and treatment. Head-and-neck CT/MR scans and organ contours from the public HaN-Seg cohort (42 cases) become 3D anatomy; a synthetic lesion placed on each patient’s buccal mucosa is simulated with PhysiCell, an agent-based model extended in C++ with time-varying cisplatin delivery and checkpoint/restart, using parameters drawn from published head-and-neck cancer measurements.

    My part. Led the simulation pipeline: quality gates, coordinate-exact mapping between simulation and CT space, a cisplatin PK/PD model with a virtual IC50 assay, a sourced drug registry, and automated invariant tests. Team project; a physician-facing 3D app with a treatment timeline is in development.

    Limits. A research prototype, not a clinical tool. The cohort has no tumour labels and one scan per patient, so lesions are synthetic and growth cannot yet be validated over time. The simulation covers a thin tissue window, not the whole lesion: agent-based models do not scale to whole-tumour cell counts.

About

Education, skills and recognition

Education

PES University (RR Campus)

B.Tech, Computer Science

Aug 2023 – 2027 · Bangalore, India

8.02/10CGPA

Data structures and algorithms, Compiler design, Machine learning, Deep learning, Generative AI, Big data, Heterogeneous parallelism, Cloud computing, Cyber forensics

IIT Madras

BS in Data Science and Applications (online degree programme)

May 2023 – present · Chennai, India

8.0/10CGPA

An online degree taken concurrently with the B.Tech.

Statistics, Software engineering, Operating systems, Computer networks, SQL and databases, Java

Tools and methods

Languages

  • Python
  • C++
  • C
  • TypeScript
  • JavaScript
  • Java
  • SQL
  • Dart
  • Solidity

ML and AI

  • PyTorch
  • TensorFlow
  • scikit-learn
  • LightGBM
  • AutoGluon
  • Chronos-2
  • TimesFM
  • FinBERT
  • MLX
  • NLP
  • generative AI
  • LLMs
  • RAG
  • fine-tuning
  • prompt engineering
  • LLM orchestration
  • LangGraph
  • embeddings and reranking
  • Ollama
  • Cerebras
  • deep RL
  • DDPG and SAC
  • anomaly detection
  • PINNs

Data science

  • NumPy
  • pandas
  • SciPy
  • Matplotlib
  • Plotly

Data and infrastructure

  • BigQuery
  • ClickHouse
  • DuckDB
  • Postgres
  • pgvector
  • Qdrant
  • Neo4j
  • MongoDB
  • MySQL
  • SQLite
  • Redis
  • dbt
  • Kafka
  • Spark
  • Docker
  • Kubernetes
  • Jenkins
  • CI/CD
  • Git
  • Linux (Debian)
  • AWS (EC2, VPC)

Web

  • Next.js
  • React
  • Astro
  • Node.js
  • Express
  • REST APIs
  • FastAPI
  • Flask
  • Django REST Framework
  • Spring Boot
  • Prisma
  • Zod
  • Tailwind
  • Nodemailer

Mobile and embedded

  • Flutter
  • Firebase
  • Android Studio
  • Arduino

CS foundations

  • data structures and algorithms
  • object-oriented programming
  • operating systems
  • computer networks
  • compiler design
  • digital design and microprocessors
  • software engineering

Methods

  • point-in-time backtesting
  • adversarial data audits
  • empirical-Bayes shrinkage
  • held-out calibration
  • claims-lag completion (chain-ladder)
  • metamorphic testing
  • retrieval evaluation (MRR, recall@k)
  • bootstrap confidence intervals
  • alerting pipelines

Security

  • malware analysis
  • DDoS simulation (isolated lab)
  • Wireshark
  • nmap
  • REMnux
  • capa
  • Cutter

Recognition

  1. 2025IMC Prosperity 4: passed Round 2
  2. 2024 – 2025CNR Rao Merit Scholarship, PES University
  3. 2024Featured on the IIT Madras official YouTube channel
  4. 2023JEE Main rank 7,786 · KCET rank 1,258 · VITEEE rank 2,504
  5. 2023CBSE Class XII: 93.6% (PCMC)
  6. 2021CBSE Class X: 97.6% · NTSE score 149 (general cutoff 159)
  7. 2018 – 2021Model United Nations: three years across committees, including Crisis
  8. SchoolModel Parliament of India, Bangalore: runner-up
  9. SchoolSTEM competition: second rank
  10. SchoolNIE Times of India National Aptitude Test: first rank
  11. SchoolICAS Mathematics: Distinction
  12. SchoolSanskrit Olympiad, all-India rank 56

Certifications

  • TCM Security, Malware AnalysisPractical malware reverse-engineering
  • AWS course completions: EC2, VPCCloud compute and virtual networking
  • Generative AI, PESU IO4-week course on NLP, sequence models, fine-tuning and LLMs
  • Flutter Primer, PESU IOCross-platform app development with Flutter and Firebase
  • IIT Madras, Foundation Level CertificateBS in Data Science programme

Clubs, leadership and volunteering

Jarvis Club

Head of the Blockchain Department

IIT Madras · Aug 2023 – May 2024

Ran workshops and hackathons on Solidity, smart contracts and Web3 front ends, and mentored students building tamper-evident records and audit trails.

Amartya Club

Member, Finance Cohort '23

IIT Madras · Dec 2023 – Mar 2024

Fundamental, technical and quantitative analysis: Piotroski, Altman, Ohlson and Beneish scores, and a technical-analysis strategy in Python.

Nexus PESU

Technical member, app development

PES University · Mar 2024 – present

Flutter and Firebase app development through hackathons, workshops and community builds.

G Cube, game development club

Sponsorship core

PES University · Feb 2025 – present

Partnerships and sponsorships for game jams and industry events.

TEDx PESU

Sponsorship member

PES University · Feb 2024 – Nov 2025

Onboarded sponsors and maintained the relationships through the event cycle.

HAYA Thor

Powertrain department

PES University · ESVC '25 · Dec 2024 – Mar 2025

SolidWorks modelling, gearbox assembly and electric-motor work for an electric solar vehicle competition.

Google Developer Student Club

Member

Bangalore · 2023 – 2024

Study jams, workshops and hackathons on full-stack, Android and cloud development.

Volunteering

Volunteer

Global Cancer Concern India · Narayana Hrudayalaya Hospitals · Ongoing

Volunteer work with a cancer-care charity and a hospital network.

Glossary

Point-in-time
Every input to a forecast was knowable at the moment the forecast claims to have been made, proven by an audit that can fail, not asserted.
Coverage
The share of the population a method can reach, not the share of what it reached that it got right. Quoting yield as coverage is the easiest way to oversell a product.

Contact

Let's buildsomething meaningful

I'm open to research collaborations, founding-team conversations, internships, and good old nerdy discussions about AI, robotics, and markets.