Candidate Snapshot
Rajas Bhingarde
Dombivli, Maharashtra
Open to ML, Applied AI, and Data Science roles
Early-career ML engineer applying AI to technical systems shaped by real-world constraints.
Electrical engineering graduate transitioning into AI and data science, with work across forecasting, NLP, automation, cloud AI, and sensor-informed research systems. I focus on useful models, measurable outcomes, and engineering decisions that hold up outside notebooks.
ML Engineer
Applied AI Engineer
Data Scientist
Fast Proof
The fastest reasons to move this candidate forward.
Featured Work
Flagship projects structured around problem, contribution, constraints, and evidence.
Quick scan Outcome, role, workflow, evidence
Technical review Decisions, constraints, evaluation
Case 01 / Research system
AI Smart Cold Storage
An IoT and ML research system combining solar forecasting, spoilage detection, and maintenance guidance for more energy-efficient cold storage.
- Outcome
- Projected 20-30% power reduction
- Ownership
- AI pipeline lead on a 3-person team
Evidence Trace
Follow information from operating context to validated claim.
Gas-sensor readings, vision inputs, solar conditions, and compressor dataplates define the documented operating context.
Open technical review Close technical review
Technical Review
Reasoning, limits, and supporting evidence
- Problem
- Cold-storage operations face power consumption pressure and spoilage risk, especially where energy use and environmental control need tighter decision support.
- Contribution
- Designed the AI pipeline, including LSTM-based solar forecasting, machine-learning logic for spoilage analysis, and the RAG-assisted maintenance guidance layer.
- Constraints
- Research-stage work required cross-disciplinary coordination, hardware coupling, and explicit separation between projected benefit and deployment evidence.
- Evaluation
- Simulation and compressor-dataplate validation support projected 20-30% power reduction potential. External evidence includes a national competition win, grant, and design registration.
Engineering Decisions
Choices grounded in documented project scope
- Selected approach
- LSTM forecasting supports pre-cooling, a gas-sensor and ResNet50 sequence handles two-stage spoilage detection, and RAG supports explainable maintenance guidance.
- Alternatives in scope
- The documented two-stage design avoids treating sensor readings or vision as a single spoilage signal. No broader architecture comparison is claimed.
- Decision constraints
- Hardware coupling, noisy physical signals, and research-stage validation favored modular outputs that could be assessed separately.
- What the work showed
- System-level benefit needs separate validation from model behavior; the power reduction remains explicitly projected.
Evidence & links
Case 02 / NLP classification
NLP Spam Detection System
A text-classification pipeline benchmarking Word2Vec and GloVe embeddings across five classifiers on more than 10,000 messages.
- Outcome
- 93% accuracy, about 8 points above baseline
- Ownership
- End-to-end individual project
Evidence Trace
Follow the benchmark from dataset to evaluation.
The benchmark uses more than 10,000 noisy, everyday text messages rather than idealized examples.
Open technical reviewClose technical review
Technical Review
Reasoning, limits, and supporting evidence
- Problem
- Spam detection needs robust classification under noisy, everyday text rather than idealized examples.
- Contribution
- Built the end-to-end pipeline, prepared text data, compared Word2Vec and GloVe embeddings, and evaluated five classifiers against a TF-IDF baseline.
- Constraints
- The comparison needed consistent preprocessing, interpretable baselines, and enough evaluation structure to avoid presenting one-off tuning as methodology.
- Evaluation
- Reached 93% accuracy, roughly 8 percentage points above TF-IDF. The repository includes precision, recall, F1, AUC, and a model comparison table.
Engineering Decisions
Choices grounded in documented project scope
- Selected approach
- Compared Word2Vec and GloVe representations across five classifiers through one consistent preprocessing and evaluation pipeline.
- Alternatives in scope
- TF-IDF served as an explicit baseline rather than an unmeasured alternative, making embedding-based improvement easier to assess.
- Decision constraints
- Noisy text and fair comparison favored repeatable preparation and the same evaluation structure across models.
- What the work showed
- A headline accuracy is more credible when the baseline, dataset scale, and multiple evaluation measures remain visible.
Evidence & links
Case 03 / Data automation
Government Energy Data Automation
A flexible extraction workflow converting inconsistent government energy reports into usable inputs for downstream forecasting.
- Outcome
- 50+ reports, weeks of parsing removed
- Ownership
- Parser and extraction workflow owner
Evidence Trace
Follow inconsistent reports into usable data.
More than 50 government energy reports arrived with inconsistent tables, embedded text, and image-heavy layouts.
Open technical reviewClose technical review
Technical Review
Reasoning, limits, and supporting evidence
- Problem
- Manual extraction from public power-sector documents was slow and error-prone because reports varied across tables, embedded text, and image-heavy formats.
- Contribution
- Reverse-engineered report patterns and built a BeautifulSoup-based extraction pipeline across more than 50 documents.
- Constraints
- Input quality and layouts varied by source, while the output needed to remain useful for downstream forecasting work.
- Evaluation
- The repeatable workflow removed weeks of manual parsing and improved dataset-preparation reliability. No deployment or public repository is claimed.
Engineering Decisions
Choices grounded in documented project scope
- Selected approach
- Reverse-engineered recurring patterns and built a flexible BeautifulSoup parser handling three documented report types.
- Alternatives in scope
- A single rigid extraction path would not cover the documented mix of tables, embedded text, and image-heavy formats.
- Decision constraints
- Source inconsistency and forecasting-ready output made adaptable parsing more important than optimizing for one layout.
- What the work showed
- Handling source variation is part of operational data engineering rather than a minor preprocessing detail.
Evidence & links
No public project artifact listed. Evidence is summarized in the resume.
Technical Judgment
How I approach constrained, messy, real-world ML work.
I work from data quality, operating context, and evaluation discipline before model novelty. The goal is not just building something impressive in isolation, but building something credible under practical constraints.
01
Start from operating conditions
I scope technical work around what the system has to tolerate: noisy inputs, limited data consistency, hardware coupling, or real process constraints.
02
Separate claims from evidence quality
I distinguish between prototype, simulated, benchmarked, and applied results so the work stays honest about what was proven.
03
Show personal ownership clearly
For collaborative work, I isolate what I designed or implemented so reviewers can judge my contribution without guessing.
04
Prefer evaluation over hype
I treat baselines, measurable outcomes, and practical usefulness as more important than fashionable models without a clear reason.
Experience
Internships framed by scope, ownership, and result.
AI & Prompt Engineering Intern
Integrated Mistral through Ollama into an application workflow, built a structured response parser for automated task routing, and reduced manual review by an estimated 3-4 hours per day.
AI Azure Intern
Built a flexible scraping and parsing pipeline for more than 50 inconsistent government energy reports, replacing slow manual extraction and improving downstream forecasting readiness.
Data Science Intern
Developed an NLP classification workflow using Word2Vec and GloVe, with visual analysis and model comparison on unstructured text datasets.
Evidence
Recognition, credentials, IP, and recruiter-ready artifacts.
Award · 2026
Tantragyan national-level hardware winner
Won 1st Prize in the Hardware Category for the AI Smart Cold Storage System.
Grant
ISHRAE research grant
Secured a ₹40,000 grant supporting cold-storage research and validation.
Intellectual Property
Registered design publication
Published design registration for the AI Smart Cold Storage System under Design No. 496644-001.
Credential · 2025
Google Cloud AI / ML skill badges
Completed eight badges covering Gemini API, Vertex AI, multimodal RAG, and ML engineering preparation.
Capabilities In Context
Capabilities tied to work, not listed as isolated tools.
Forecasting under constraints
Time-series reasoning tied to system use
Used LSTM-based solar forecasting within an energy-efficiency concept where prediction informed operational decisions.
NLP classification
Benchmark-driven text modeling
Structured NLP work around baseline comparison, embedding choices, and measurable improvement.
AI + sensor workflows
Machine learning adjacent to physical systems
Worked on AI concepts that interact with hardware, sensing, and environmental constraints.
Operational automation
Reducing manual work with structured pipelines
Built report-extraction and LLM-assisted routing workflows where consistency mattered more than novelty.
About
Engineering foundation first, curiosity disciplined by evidence.
My electrical engineering background shaped how I think about systems: signals, operating limits, instrumentation, and why real environments rarely behave like clean datasets. That foundation pulled me toward applied AI rather than purely abstract model work.
I am especially interested in forecasting, NLP pipelines, automation, and sensor-informed systems. Astrophotography influences my mindset more than my aesthetics; it taught me patience, calibration, and respect for noisy signals.
Education
B.E. Electrical Engineering
Lokmanya Tilak College of Engineering, with a transition into AI and data science through applied project work.
Working style
Methodical and evidence-led
I prefer clear evaluation criteria, scoped contributions, and systems thinking over inflated technical claims.
Contact / Handoff
Ready for recruiter review and deeper technical discussion.
Best fit for ML engineering, applied AI, and data science roles involving technical systems, forecasting, NLP, or operational automation.
- rajasbhingarde@gmail.com
- +91 9372869511
- Dombivli, Maharashtra, India