Deepak Karkala

SYSTEMS BUILT:

Arbiter

• AI governance and decision-intelligence platform for engineering teams
• CLI-first workflows, audit trails, confidence scoring, connector-backed context, and async review loops
• Currently in beta testing

Creative Collab OS

• Beta creative AI studio for stand-up, songs, comics, sitcoms, and memes
• Workflow graphs, direction checkpoints, taste-memory profiles, streaming outputs, and multimodal artifact generation
• Currently in beta testing

BatSwing

• PWA + computer-vision cricket coaching product for one-phone side-on batting capture
• Pose/phase analysis, WhatsApp-ready academy reports, and consent-aware player media handling
• Processing reports weekly while tuning the pipeline

Tech Abstractions

• AI/ML interview-prep platform focused on production reasoning, system design, LLM infrastructure, and applied practice
• Structured learning paths, workbook-style exercises, expert walkthroughs, and course-linked practice
• A few dozen users on the waitlist

Dev Tools and Illustrated Explainers

• Dev tools, plugins for coding agents, and repeatable workflows for Agentic-AI engineering and Production MLOps
• Illustrated explainers covering Coding agents, LLM inference, RLHF, Statistics-for-MLOps, and related AI infrastructure topics

Fine-tuning Projects

• Sitcom Scriptwriter: small language model fine-tuning for screenplay-style generation with custom datasets and evaluator workflows
• Kannada Physics Tutor: local-language physics tutoring with staged fine-tuning, evaluation harnesses, and RAG/evaluator components

Tech Stack: LLMs, Agentic AI, MLOps, Fine-tuning, Evaluation, Coding Agents, Product Engineering

SYSTEMS BUILT:

CLV Prediction

• Segmented models (early/established); RMSE 15-20% improvement, Gini 0.47→0.62
• Diagnosed/resolved feedback loop via cross-functional experiment
• Weekly batch scoring 300K+ profiles with segment monitoring
• XGBoost, SageMaker, Glue, S3

Purchase Intent Scoring

• Sub-second scoring 60K+ daily predictions (Feast + Redis); 50K+ events/hour via Spark
• Resolved calibration breakdown; segment-wise analysis + post-hoc calibration restored trust
• A/B testing delivered 5% uplift; 40% p99 latency reduction
• LightGBM, Feast, Redis, Kinesis, SageMaker

RAG Search & Discovery

• Hybrid retrieval (embeddings + BM25) + re-ranking; LLM-driven golden dataset
• 3M queries/month, p99 <500ms, EUR 0.001 per query; 4% search-to-purchase uplift
• OpenAI, Pinecone, GPT-4, fine-tuned re-ranker

Review Summarization

• Mistral-7B on 100K+ reviews (LoRA + vLLM); partnered with legal for governance
• Automated quality checks (hallucination/toxicity); 60% cost vs. GPT-4
• Mistral-7B, vLLM, LoRA, SageMaker

Tech Stack: Python, LightGBM, XGBoost | AWS (SageMaker, Glue, Kinesis, Lambda) | Terraform | Feast + Redis | OpenAI, Mistral

Spiticart - E-commerce platform for traditional Indian handicrafts

• Direct-to-consumer model supporting rural artisans
• Products: Shantiniketan leather, Channapatna toys, Agra marble carvings
• Mission: Preserving traditional crafts and improving artisan livelihoods

Rumi Schools - Digital-first rural education initiative

• Designed comprehensive curriculum spanning academics, co-curricular activities (tech, arts, sports), and life skills
• Developed prototype leveraging digital technologies and remote mentorship for scalable impact
• Part of larger "Model Village" vision for sustainable rural communities

.

**SYSTEMS BUILT**:

Predictive Maintenance (Anomaly Detection + HITL)

• Deployed anomaly detection + alert triage for heating systems
• Diagnosed alert fatigue issue; partnered with maintenance teams to redesign as prioritization tool vs. automated detector
• Evolved from unsupervised (residuals + LOF) to supervised as labels grew via technician human-in-the-loop
• 75% precision@50 for high-priority alerts; restored stakeholder trust
• Python, scikit-learn, time-series analytics, MLOps pipelines

Energy Forecasting (Smart Energy Advisor)

• 24-hour ahead demand forecasting (XGBoost + weather, lag, rolling windows, holiday features)
• Powered resident-facing recommendations; 10pp solar self-consumption increase
• Tiered cold-start strategy for new buildings: physics-informed heuristics → archetype models → individualized forecasts
• Walk-forward validation, baselines (ARIMA/Prophet), drift monitoring
• Python, XGBoost, time-series pipelines

Tech Stack: Python, scikit-learn, XGBoost, ARIMA | PostgreSQL | Pandas | Docker

Thesis: Data Analysis and Anomaly Detection in Buildings Using Sensor Data

• Developed unsupervised learning methods for IoT anomaly detection
• Applied to real building management systems
• Foundation for later production IoT ML work at eSMART Technologies

Relevant Coursework:

• Machine Learning
• Applied Data Analysis
• Computer Vision
• Natural Language Processing
• Algorithms

This degree provided rigorous foundation in ML theory and practical application that enabled rapid transition to production ML engineering.

RESEARCH CONTRIBUTIONS

Traffic Surveillance Object Detection

• Fine-tuned YOLOv2 for tandem motorbike detection in traffic surveillance video
• Tuned anchor boxes/heads with targeted augmentation to surface candidate "tandem near pedestrian" incident windows for human review

Scene Understanding (Visual Relationship Detection)

• Built PyTorch relationship detector with scene graph outputs (subject-predicate-object)
• Engineered spatial features and explored graph/message-passing + translation embedding formulations
• Improved predicate classification robustness for downstream incident reasoning

Tech Stack: Python, PyTorch, YOLOv2

KEY ACHIEVEMENTS

• Developed and validated WCDMA uplink receiver algorithms (Path Searcher, RAKE, Multiuser Detection)
• Achieved 3GPP NodeB receiver conformance (BER <0.001) across multipath/interference channel conditions
• Reduced storage and computational complexity in baseband receivers through novel algorithm design

PATENTS

• US-9602240-B1 (Granted Patent, 2017): Symbol-level interference cancellation for WCDMA uplink receivers. Reuses soft-bit estimates across iterative decoding (serial or parallel IC) to cut memory/compute while maintaining detection accuracy in multiuser scenarios.

• US-20160365991-A1 (Published Application, 2016): Two-stage channel estimation for WCDMA: a coarse search finds regions of interest in the uncertainty window, then a fine search runs only in those regions. This reduces correlation operations and hardware complexity while preserving estimation accuracy.

This role provided deep foundation in algorithms, optimization, and systems-level engineering that later translated to ML systems work.

Tech Stack: MATLAB, Octave, Signal Processing, WCDMA, LTE

RESEARCH CONTRIBUTIONS

• Developed a data-resolution matrix method (from the sensitivity/Jacobian model + regularization) to identify independent measurements
• Used diagonal vs. off-diagonal structure to select informative measurements and reduce data-collection time
• Achieved ~20% more independent measurements than singular-value selection while preserving reconstruction quality

Tech Stack: MATLAB, Optimization algorithms, Inverse problems

Strong foundation in mathematics, signal processing, and algorithms that provided the technical base for later work in ML and wireless communications.