- Education: BSc (Hons) Artificial Intelligence & Data Science undergraduate at IIT Sri Lanka (affiliated with Robert Gordon University)
- Industry Experience: Former AI Engineering Intern at DecodeLabs — delivered end-to-end applied ML, NLP recommender, guardrail, and OpenCV/OCR pipelines
- Engineering Focus: Multimodal OCR forensics & claim auditing, adversarial tool-agent evaluation, hybrid retrieval (RAG, FActScore NLI, G-Eval), streaming ML observability, and open-source ML tooling
View Detailed Methodology & Verified Metrics
| Repository | Scope & Engineering Details | Verified Results & Highlights |
|---|---|---|
veridoc-forensics |
Multimodal document auditor combining OpenCV skew/blur restoration, JPEG Q90 ELA + MAD noise tamper heatmaps, Tesseract OCR, hybrid BM25/TF-IDF claim verification, and FastAPI Studio UI | Restores degraded OCR char accuracy from 0.0% → 99.7% (66.7% → 99.9% mean across suite), 100.0% claim verdict & tamper accuracy, 16 passing pytest tests |
traceguard-eval |
Deterministic multi-step tool-agent sandbox, adversarial prompt-injection & HTTP/JSON fault fuzzer, schema/budget guards, and FastAPI Trace Replay UI | TraceGuardHardenedAgent scores 100.00 vs 63.33 (SchemaRetryAgent) & 39.58 (NaiveAgent) on benchmark (100.00 vs 61.00 vs 37.69 on 30-run fuzz); 17 passing pytest tests |
CM3604_CW |
Automated fact-checking of LLM outputs against BM25 evidence passages with grouped CV ablations & error analysis | Majority baseline 50.68% (112/221) → Lexical overlap 84.16% (186/221) → Fine-tuned DeBERTa NLI 88.24% (195/221) |
DecodeLabs-Internship |
Four applied AI capstones spanning input guardrails, customer classification, TF-IDF skill-gap recommendations, and document OCR | Combines OpenCV perspective/contrast normalization with Tesseract extraction across end-to-end runnable modules |
seniruvitharanage/domain-ai |
Contributed priority-scoring ML sub-pipeline merging TF-IDF developer issue text with repository & workflow metadata | Gradient Boosting model trained on 9,120 samples achieving 96.97% accuracy & weighted F1 on 2,280 test cases |
CM2604-Churn-Classifier |
Telco customer churn prediction comparing hyperparameter-tuned Decision Tree and Dropout/EarlyStopping Keras MLP | Strict train-only SMOTE oversampling to prevent data leakage under ROC-AUC and generalization-gap evaluation |
| Repository | Upstream Project | Engineering Contributions & Verification |
|---|---|---|
Thush099/chonkie |
feyninc/chonkie |
Fixed OverlapRefinery._split_text & _recursive_overlap infinite-loop / whitespace chunk boundary bugs (fix/overlap-refinery-recursive-whitespace, upstream #681); 89 refinery & recursive chunker tests pass (4 from_recipe tests deselected without huggingface_hub) |
Thush099/cleanlab |
cleanlab/cleanlab |
Fixed IndexError bounds ordering in token_classification.rank.issues_from_scores (upstream #1338) and single-class K == 1 ValueError in rank.get_normalized_margin_for_each_label (upstream #1356) on fix/rank-boundary-guards; 124 upstream/regression tests pass across rank, token classification, util & validation modules |
Thush099/hiring-agent |
interviewstreet/hiring-agent |
Resolved scoring clamping, JSON extraction & defensive transform bugs and built calibration_audit.py for offline rubric integrity (100.0% tier accuracy, 97.2% evidence grounding) and counterfactual fairness auditing (0.00 delta, 1.000 Kendall's τ) on feat/offline-rubric-calibration-audit; 13 passing tests |
A suite of standalone, modular Python packages covering retrieval, LLM faithfulness, production drift monitoring, differential privacy, causal inference, and multimodal attention — each shipped with runnable CLIs and automated pytest suites.
LLM Systems & Evaluation
|
ML Observability & Drift Monitoring
|
Data Contracts & Federated Privacy
|
Causal & Multimodal Learning
|
Self-hosted SVG telemetry and project showcase cards generated directly from GitHub GraphQL via scripts/generate_profile_cards.py.


