Parth Malasne — Portfolio

Projects

Certified

Self-directed project

Predictive Maintenance: Industrial ML for Fault Detection

Condition-Monitoring Engineer · AI / ML · October 2026

92/ 100

A ProoV case study · educational project, not employment

Built an end-to-end rolling-bearing fault detector from vibration data: engineered RMS, kurtosis, crest factor, and envelope-spectrum defect-frequency features; trained a depth-capped decision tree on training bearings only; and evaluated it on held-out bearings with catch-rate and false-alarm metrics. The candidate also localized the fault to the outer race via BPFO and justified an alarm threshold with a clear cost-based maintenance memo.

Graded against

Feature engineering (signal to honest numbers)25%
Held-out evaluation (honest measurement)25%
Cost-based threshold decision + memo25%
Compact model + embedded-budget justification15%
Plain-language explanation10%

Passed · pass mark 60/100

What stood out6

Correctly tied the dominant envelope peak to BPFO and the outer race fault location
Chose kurtosis as the time-domain feature that jumps most on impulsive faults
Reported a compact tree with stated depth 2 and 5 nodes, aligned to the embedded budget
Defended the 0.40 threshold using explicit €20,000 miss cost versus €500 false-alarm cost
Leakage-Free Feature Engineering
Honest Held-Out Evaluation

The work I submitted19 tasks

11 tasks · 5.9k characters · Python, JavaScript · 8 written answers · 383 words

Initial Prediction62 characters
Load Plot Signal CodePython · 1.8k characters
Load Plot Signal205 characters
Example Rms WorkedPython · 697 characters
Time Features143 characters
Freq FeaturesPython · 775 characters
Fault Location239 characters
Train Classifier363 characters
Evaluation Heldout288 characters
Threshold Decision334 characters
Decision MemoJavaScript · 1.0k characters
Initial Prediction9 words
Load Plot Signal Code163 words
Load Plot Signal15 words
Example Rms Worked72 words
Time Features10 words
Freq Features69 words
Fault Location14 words
Train Classifier31 words

Summarised on purpose — the submitted code and writing stay private so this page cannot be reused as an answer key. The full submission sits behind the verified certificate.

Verified certificateTamper-proof · issued by ProoV

Certified

Self-directed project

Energy and Renewable: Machine Learning for the Power Grid

Junior Machine Learning Engineer · AI / ML · October 2026

92/ 100

A ProoV case study · educational project, not employment

Built a leakage-aware turbine yield predictor end to end: framed TEY as the target, excluded downstream CO/NOX emissions, trained a LinearRegression baseline and a RandomForest model, and validated on a later held-out period. The final model achieved 1.53 MWh MAE, 1.94 RMSE, and 0.983 R² on held-out data, with a model card that clearly states intended use, limits, and monitoring triggers.

Graded against

Build and validate the model correctly35%
Optimize and evaluate honestly25%
Write the Model Card (clarity + honesty)25%
Explain it to a non-expert15%

Passed · pass mark 60/100

What stood out6

Correctly excluded CO and NOX as downstream leakage features and explained why they are unavailable at prediction time.
Used a later-year holdout split and reported held-out MAE, RMSE, and R² rather than relying on training performance.
Included a residual read and acknowledged that strong performance on one plant does not guarantee generalization to other turbines.
Wrote a concise manager-facing summary that correctly states the prediction target, accuracy, trust basis, and failure mode.
Leakage-Free Feature Selection
Time-Based Holdout Evaluation

The work I submitted20 tasks

12 tasks · 13 lines · Python, JavaScript · 8 written answers · 416 words

Problem Frame214 characters
Leakage Flags113 characters
Train Validate BaselinePython · 1.4k characters
Train Validate Forest943 characters
Train Validate124 characters
Tune Evaluate SetupPython · 1.5k characters
Tune Evaluate ResidualsPython · 2 lines
Tune Evaluate132 characters
Residual Read345 characters
Drift Monitor93 characters
Model Card1.0k characters
Teach BackJavaScript · 726 characters
Problem Frame16 words
Leakage Flags1 word
Train Validate Baseline141 words
Train Validate Forest76 words
Train Validate1 word
Tune Evaluate Setup110 words
Tune Evaluate Residuals70 words
Tune Evaluate1 word

Summarised on purpose — the submitted code and writing stay private so this page cannot be reused as an answer key. The full submission sits behind the verified certificate.

Verified certificateTamper-proof · issued by ProoV

Certified

Self-directed project

AI-Assisted Code Modernization with IBM

Junior Software Engineer · Software Engineering · October 2026

96/ 100

Based on a public IBM case study · not affiliated with IBM

Built and verified an end-to-end modernization of a legacy fleet-service repo: fixed the wear calculation, handled missing service readings safely, corrected the nightly report’s average and crash path, and found a hidden mileage-conversion bug in the helpers. Also added the missing test, produced a data-driven breakdown-risk analysis from fleet_history.csv, and documented the review process in NOTES.md with clear acceptance checks.

Graded against

Fix and extend the repo with the agent35%
Diagnose the bug20%
Direct and audit the AI agent25%
Explain it like an engineer20%

Passed · pass mark 60/100

What stood out6

Correctly identified the integer-division wear bug and explained the real consequence: nearly-worn cars could read as 0% and never be flagged.
Caught the quiet helper-module conversion bug in fleet_utils.py, which no test covered, showing a genuine sweep beyond the obvious failures.
Wrote a concrete NOTES.md that documents what the agent got wrong and how the fixes were verified with verify.py and pytest.
Built an analysis that follows the data: it explicitly rejects odometer_km and age_years as primary predictors and emphasizes km_since_service, load_factor, and avg_daily_km.
Leakage-Free Bug Diagnosis
Verified AI-Agent Auditing

The deliverable

View the codeparthmalasne888-crypto / proov-km-waechter-fix

The work I submitted20 tasks

12 tasks · 2.0k characters · 8 written answers · 114 words

Division Predict70 characters
Diagnose Line130 characters
Diagnose Why138 characters
Scope Prompt66 characters
Audit Findings158 characters
Audit Why259 characters
Acceptance Checklist263 characters
Bob Ready60 characters
Template Repo Click82 characters
Verify Report71 characters
Exec Pitch561 characters
Bob Bonus182 characters
Division Predict1 word
Diagnose Line10 words
Diagnose Why22 words
Scope Prompt1 word
Audit Findings1 word
Audit Why39 words
Acceptance Checklist39 words
Bob Ready1 word

Summarised on purpose — the submitted code and writing stay private so this page cannot be reused as an answer key. The full submission sits behind the verified certificate.

Verified certificateTamper-proof · issued by ProoV

Certified

Self-directed project

Machine Learning for Automotive Safety

Perception Validation Engineer (ADAS Safety Sign-off) · AI / ML · July 2026

92/ 100

A ProoV case study · educational project, not employment

Built a perception validation dossier end to end for a synthetic Aurora-7 detector: defined detections with an IoU >= 0.5 rule, scored precision/recall and VRU false-negative rate from a synthetic table, and sliced recall by night, occlusion, and rarity to expose the worst safety gaps. The final recommendation was evidence-based and conditional, with SOTIF-style mitigations tied to the specific night-and-occlusion pedestrian risk and a correct comparison to the Uber Tempe crash mechanism.

Graded against

Detection metrics correctness (precision, recall, VRU false-negative rate)30%
Failure-mode analysis & per-condition rigour25%
SOTIF safety case quality (ISO 21448 reasoning)25%
Ship verdict — evidence-based & internally consistent20%

Passed · pass mark 60/100

What stood out6

Computed and carried forward the correct IoU-based detection rule, including the 0.5 match threshold and unmatched-real-object false negative definition.
Used VRU recall and VRU false-negative rate as the safety lens, rather than relying on aggregate accuracy alone.
Exposed the dangerous slices clearly: night recall 0.39, occluded recall 0.45, and worst-case night+occluded recall 0.27.
Wrote an internally consistent GO-WITH-CONDITIONS verdict that cites the dossier’s own metrics and mitigation path.
Leakage-Free Detection Definition
Honest Safety Metric Evaluation

The work I submitted20 tasks

12 tasks · 14.9k characters · Python · 8 written answers · 817 words

Sensor Match553 characters
Example Iou Demo1.5k characters
Task Definition Code1.1k characters
Task Definition377 characters
Example Metrics DemoPython · 2.2k characters
Scored Metrics RunPython · 2.3k characters
Scored Metrics351 characters
Realcase Analysis830 characters
Teachback741 characters
Ttcs Example SlicePython · 1.3k characters
Ship Verdict1.5k characters
Sotif Case2.2k characters
Sensor Match27 words
Example Iou Demo169 words
Task Definition Code149 words
Task Definition42 words
Example Metrics Demo154 words
Scored Metrics Run143 words
Scored Metrics18 words
Realcase Analysis115 words

Summarised on purpose — the submitted code and writing stay private so this page cannot be reused as an answer key. The full submission sits behind the verified certificate.

Verified certificateTamper-proof · issued by ProoV